Who Answers for the Machine? No. 2 of 3. Why agent identity is not accountability, and what Singapore is really building.
In Parliament this month a Singapore Senior Minister of State gave the clearest one-line statement of AI accountability I have heard from any government: “Agents can act for us, but they cannot answer for us.” Singapore is now exploring how to give every agent a name that can be checked across the internet. This essay argues that the name is necessary, that it is not enough, and that Singapore’s own documents already point to the layer that is missing: proof of what the agent actually did.
What This Updates
This essay stands on its own, but it extends five earlier Genesis pieces, and readers of those deserve to know what is new.
- The Last Human Job Is Answerability (22 October) argued that what remains ours is not empathy but the willingness to answer for what we do. Jasmin Lau’s line in Parliament is the policy version of that claim. Here I test it against the machinery being built to deliver it, and sharpen it: identity makes someone blameable; only evidence makes them answerable.
- AI Agent Assurance: A Board’s Guide to the Governance Frameworks (6 November) compared the governance frameworks and put “identity before rights” first among its six components. This essay adds the layer that comparison did not cover: the identity protocols themselves, eleven of them, read in their own words.
- Stopping Deepfake Fraud at Runtime and Every AI Instruction Should Carry Proof (Receipts, Please, Parts IV and V) argued that every high-value instruction should carry its own proof at runtime. Here I show why an identity credential is not that proof, and I have now read the MAS SAFR paper itself, which says the same thing in one sentence.
- Six Regimes, One Agent: The November Update (16 November) mapped the regulators. A reader, Baris Tuncertan, asked how banks should prepare for the incident clock that is coming. Section XIII is my answer, from the identity side.
The Two-Minute Version
- On 4 November, answering an adjournment motion by Tin Pei Ling, who proposed a Singpass-like national registry for AI agents, Senior Minister of State Jasmin Lau said IMDA is working with GoDaddy, Infoblox, Cloudflare and Google to explore adapting the domain name system so agents can be identified and verified across platforms. She did not commit to a registry. Singapore is exploring, not adopting.
- Her sentence, “Agents can act for us, but they cannot answer for us”, is the best available statement of what I call the registration theory of accountability: if every agent can be traced to a responsible human, accountability follows. It is half right.
- I read eleven identity and authorisation proposals, from DNS-based discovery and IETF drafts to Visa, Mastercard, Google, Microsoft and Okta. Every one of them, in its own words, proves who is asking or what was permitted. None claims to prove what the agent actually did. Several say so explicitly. The most important sentence I found is in MAS’s own SAFR paper: agent identity “should not be understood as a mechanism for trusting the agent’s reasoning or output.”
- Six failure modes pass every identity check: duplicate execution, claims that do not match actions, drift inside a granted scope, prompt injection, broken delegation chains and the gap between revoking an agent and its pending actions stopping.
- Accountability needs three layers: who the agent is, what it may do, and what it actually did. The world is building the first two. Singapore’s own documents, the agentic governance framework, the agent testing starter kit, the GovHarness speech and SAFR, already ask for the third.
- Singapore should keep exploring identity, and in parallel define the record: a common evidence format for agent actions, independent of the agent, that a registry entry can point to. As ASEAN chair in 2027, it can make that format travel.
- On Monday morning: CIOs should map which actions in their estate are proven, not just permitted; CROs should ask for the receipt from the last time a gate said no; regulators should test the record, not just the registration; boards should ask who would answer, with what evidence, if an agent did something twice.
I. The Look-Alike Seller
Picture the demonstration hall at GovWare, during Singapore International Cyber Week in October. On the screen, a traveller wants a flight from Madrid to Singapore. She does not book it herself. Her agent does, running on Amazon Web Services. The seller is also an agent, running on Cloudflare Workers. The two have never met. A third agent, running on Google, acts as the mandate authority and verifier.
The traveller approves the purchase with a passkey. That produces a signed mandate, in Google’s Agent Payments Protocol, stating what her agent may buy, from whom, and up to what amount. Her agent then looks for a seller through the domain name system, the internet’s oldest directory. It finds one, checks the seller’s identity against a public log, verifies its trust score, and buys the ticket.
Then the demonstrators run it again with a twist. A look-alike seller appears, carrying forged credentials. The buyer agent checks, the checks fail, and the transaction never happens. The fake is blocked before a cent moves.
The project, called the Chain of Trust, was built by GoDaddy, Infoblox, Google and Cloudflare with Singapore’s Infocomm Media Development Authority, and it draws on the agentic AI governance framework I co-authored. It is a demonstration and an open reference architecture, not a deployed national system. Lee Wan Sie, IMDA’s cluster director for AI governance and safety, framed its purpose simply: “Robust agent identity will be a key part of that.” Jared Sine of GoDaddy gave the engineering philosophy: “We don’t need to rebuild the internet for AI agents.” Wei Chen of Infoblox described the prize: businesses that can independently verify who they are dealing with, without one platform deciding who is trusted.
It is a good demonstration. I want to be clear about that before I complicate it, because the rest of this essay could be misread as scepticism about identity. It is not. A world in which a buyer agent can tell a real airline from a forged one, using infrastructure no single company owns, is better than the world we have now, in which that question is mostly answered by whoever runs the platform.
But watch where the demonstration stops. The forged seller is caught because its credentials fail. Now run the third version, the one nobody demonstrates. The seller is real. Its credentials are perfect. The buyer’s mandate is valid, signed by a passkey, limited to one ticket under a set price. Every check passes. And then something goes wrong after the check: the seller’s agent charges twice because a retry fired after a timeout; or it books the wrong date because a field was misread; or the buyer agent, mid-task, reads a booking page with hidden instructions in it and decides to add travel insurance the traveller never wanted; or the payment succeeds, the ticket is never issued, and the buyer agent reports to its owner that all is well.
None of these failures involves a forged identity. All of them involve real, verified, authorised agents doing something other than what was intended. And in every one of them, the question the traveller will ask is the one Singapore’s minister put to Parliament: who answers for this?
The Chain of Trust tells her exactly who the seller was. It does not tell her what happened. That gap, between knowing who acted and knowing what they did, is the subject of this essay.
The first essay in this series looked at a different kind of claim meeting a different kind of reality: a safety threshold in Anthropic’s prospectus becoming a securities disclosure, where the question is who answers if the disclosed claim and the actual model diverge. The thread that runs through all three essays is that accountability for AI is arriving in late 2026 through disclosure and runtime evidence, not through AI law. This one asks the same question of the agent itself. The third, Exactly Once, will go deep on the runtime layer where the answer has to be built.
II. The Best Sentence in Parliament This Year
On Wednesday 4 November, Tin Pei Ling, Member of Parliament for Marine Parade-Braddell Heights, used an adjournment motion to propose that the Government consider building a national digital identity infrastructure for AI agents. As reported by The Straits Times, she suggested an open and trusted registry against which any party could check an agent’s identity and what it is authorised to do: something akin to Singpass, the national digital identity that most Singaporeans use to log in to government and banking services, but for AI agents. I have not found a verbatim quotation from her speech in the reports I could read, so I paraphrase it here and nowhere put words in her mouth. One trade outlet, Biometric Update, described her proposal as “biometric passporting” for agents; that phrase is the outlet’s, not hers as far as I can verify, and I will not lean on it.
Jasmin Lau, Senior Minister of State for Digital Development and Information, replied. She said the Government shared Ms Tin’s ambition for Singapore to be a place where people and businesses can use AI agents with confidence, and that this requires agents’ identity, authority and accountability to be verified across different systems. She described the Government’s approach as working with industry “to find out what works in practice”, building on existing infrastructure and emerging international standards, and stepping in with common standards or infrastructure “where there are gaps that industry cannot fill on its own”. She said the answer would come from many efforts, “from recognising agents at checkouts to checking their authority before a transaction, to identifying them across platforms”. She disclosed that IMDA is working with GoDaddy, Infoblox, Cloudflare and Google to explore adapting the domain name system so that agents can be identified and verified across platforms and organisations, and that this work “will inform our assessment of whether additional common infrastructure, including a national registry, is needed”. Then she said the sentence that gives this essay its title:
“Agents can act for us, but they cannot answer for us. Our task is to make sure that behind every agent, there is always someone who does.”
I have read a great deal of government writing about AI this year. I have not read a better two sentences. They do three things that most official statements avoid. They concede that agents act, rather than pretending agents merely “assist”. They locate the problem precisely, in answering, not in acting. And they assign a task to the state: to make sure that someone stands behind every agent.
Readers of The Last Human Job Is Answerability will recognise the shape. In that essay I argued, against Kai-Fu Lee’s claim that love and empathy are what will keep us human, that the last human job is answerability: the standing to be called to account, to give reasons, to bear the cost of an outcome. I borrowed from the philosophers Gary Watson and David Shoemaker the idea that responsibility has three faces. Attributability asks whether an act expresses who you are. Answerability asks whether you can be called on to give your reasons. Accountability asks whether others may hold you to demands, with blame or gratitude, because you have regard for them. I argued that machines can increasingly perform the first, that they cannot stand in the other two, and that the law, the insurers and the unions were already rebuilding answerability case by case.
Ms Lau’s line is that argument in the mouth of a government. “They cannot answer for us” is the answerability face, stated as policy. It is a vindication I did not expect so soon, and I am grateful for it. But a vindication is also a test. If the state now accepts that someone must answer behind every agent, the next question is what that someone needs in order to answer. And here the policy, as it is currently being built, risks quietly substituting one face of responsibility for another.
The registration theory of accountability
Call it the registration theory of accountability. It runs like this. Accountability fails when an agent’s actions cannot be traced to a responsible person. So give every agent a verifiable identity, bind that identity to a human or an organisation, and keep a register. Then, when something goes wrong, the register tells you who to call. Responsibility follows the name.
It is an old and powerful idea, and much of modern economic life rests on it. Company registers, vehicle registration, the licensing of professionals, the know-your-customer rules in banking: each works by binding an actor to a name that someone in authority can look up. Ms Tin’s proposal is a direct application of it to agents. So, in part, is the IMDA work on DNS, and so are most of the commercial products that now promise to give agents identities.
The registration theory is not wrong. It is incomplete in a specific way that philosophy, again, helps to name. A register delivers attributability: this act came from that agent, which belongs to that person. It can deliver the minimum condition for accountability: there is someone to blame. What it does not deliver is answerability in the full sense, the ability of that person to give an account of what happened and why. To answer, you need to know what was done. You need the reasons, the inputs, the sequence, the authority that was checked and the outcome that followed. A registered owner who cannot reconstruct what their agent did cannot answer for it. They can only be blamed for it.
That is the sharpening I want to add to my October essay. Identity makes someone blameable. Only evidence makes them answerable. Blame without an account is not accountability. At its worst it is scapegoating: the named human pays for an outcome they could neither see nor explain. At its most common it is something duller and more corrosive: a name on a form, a sponsor field in a directory, a “Directly Responsible Individual” who learns what their agent did from the customer’s complaint.
Ms Lau’s second sentence already contains the remedy, if we read it carefully. “Behind every agent, there is always someone who does” answer. Answering is a verb about giving an account. The task she set is not only to make sure there is a name behind every agent. It is to make sure that whoever stands behind it can actually answer. That requires a record of what the agent did, kept somewhere the agent cannot edit, in a form the human can read. Without that, Singapore will have built an excellent directory of people who cannot explain what happened.
III. What Singapore Is Actually Building

It is worth being precise about what exists, because commentary on the November debate has already started to inflate it. Singapore has not adopted a national agent registry. It has not mandated DNS-based agent identity. It has not chosen a standard. Here is what the public record supports, as of 17 November.
IMDA’s DNS exploration. IMDA is working with four companies, GoDaddy, Infoblox, Cloudflare and Google, to explore adapting the domain name system so that agents can be identified and verified across platforms and organisations. The minister’s own word was “explore”, and she tied the outcome to a later assessment of whether “additional common infrastructure, including a national registry, is needed”. AsiaOne’s report of the debate went slightly further, saying such infrastructure “may eventually be needed”. Either way, it is a question, not a decision.
The Chain of Trust demonstration. Shown at GovWare in October, developed with IMDA, building on the agentic governance framework (version 1.5). It combined DNS-based discovery, OAuth, the Model Context Protocol, Google’s Agent Payments Protocol mandates, passkeys and a public log. The participating companies published a technical case study and an open reference architecture and invited registrars, identity providers, payment networks and platforms to adopt and extend it. It is a proof of concept, and a well-chosen one.
Tin Pei Ling’s registry proposal. An MP’s proposal in an adjournment motion, not government policy. Its value is that it put the question on the parliamentary record and drew a substantive ministerial reply.
GovTech’s internal agent registry. Ms Lau said GovTech is developing an internal registry to help public agencies keep track of the AI agents they deploy and the officers responsible for them. This is the registration theory applied inside government, where it is easiest to apply: one employer, one directory, clear lines of authority. It is in development.
GovHarness and the ARC framework. On 5 November, at GovTech’s STACK conference, Senior Minister of State Tan Kiat How said: “GovTech is developing GovHarness as shared infrastructure for building and operating AI agents across Government.” He described support for deployment, connections to tools, and “capabilities for monitoring and evaluation”, and said it “will also include common guardrails and evaluations”. He also described GovTech’s Agentic Risk and Capability (ARC) framework, with requirements “covering human oversight, permissions management, tool access, testing, and monitoring”, and said higher-risk actions such as “financial transactions, official communications and data deletion” need stronger controls. Three of his lines deserve to be pinned above every agent team’s desk: “Building quickly is not the same as deploying safely.” “If an agent produces an unexpected result, can we retrace its steps?” And, the one that matters most for this essay: “We need to know not only what agents are allowed to do, but whether they are doing it as intended.”
MAS SAFR. Safeguards for Agentic Finance at Runtime is a paper developed by the financial industry under MAS’s BuildFin.ai initiative, first published on 3 July and, on the evidence of the current file, now at version 1.1, with the MAS page last updated on 14 October. I have now read the paper itself, which earlier Genesis pieces described from secondary reads. It is explicit that “It does not constitute regulatory guidance or supervisory expectations.” It sets out five elements: an Agent Identity component that binds each proposed action to a registered agent and checks it against the registry before anything else; a Controls Repository holding the rules, from policy and regulation to the user’s own mandate; a Disposition Engine that evaluates each in-scope action deterministically and returns Deny, Escalate, Auto-Execute or Observe; a Governance Envelope that packages the action with its trace and context before execution; and a tamper-evident, append-only Audit Log of every decision. The acknowledged institutions are Ant International, Circle, DBS, Google, HSBC, J.P. Morgan Chase, Manulife, Mastercard, OCBC, UOB and Visa.
The agentic governance framework, version 1.5. IMDA’s Model AI Governance Framework for Agentic AI, first launched in January, updated in May and again in June. I co-authored it, which readers should weigh. On identity it says plainly: “An agent should have its own unique, cryptographically verifiable identity”, tied to a human, a supervising agent or a department. On authority it asks for authorisations that are “scoped, time- or session-bound, non-transferable”, says a user cannot grant an agent more than the user’s own permissions, and that “delegations of authority should be clearly recorded”. It flags “recursive delegation scenarios where agents spin up multiple sub-agents” as an open problem. And on runtime it asks for “automated real-time monitoring to escalate any unexpected or anomalous behaviour”, noting that “Runtime controls address this by monitoring and intervening during execution.”
The Starter Kit for testing agentic systems. Version 1.0, released in October and announced on 14 October, is guidance, not regulation. Its treatment of identity is brief, a single question: “What identities, credentials and privileges does the agent possess or inherit?” Its treatment of what happens afterwards is extensive. It asks testers to “capture the entire agentic trajectory”, says “logs should be generated or independently corroborated outside the agent’s control”, and warns, in a sentence that should be read slowly: “An agent may claim that it had completed an action without having actually done so.” It says “agentic systems should be continuously monitored and periodically re-tested”. It asks: “Are the agent’s actions reversible?”
ASEAN. Ms Lau said, per AsiaOne’s report, that Singapore intends as ASEAN chair in 2027 to strengthen regional capabilities in AI governance, testing and assurance, including adopting international standards and common testing tools for agentic AI.
Lay those eight items side by side and a pattern appears that the headlines missed. The newest, most reported item, DNS-based identity, is the one Singapore is least committed to. The items Singapore has actually written down, in frameworks, testing kits, speeches and an industry paper, spend most of their words on something else: on what the agent did, whether it was within its authority, whether the record can be trusted, and whether anyone can retrace its steps. Singapore’s identity layer is exploratory. Its evidence layer, on paper at least, is already specified. That asymmetry is the opportunity this essay is about.
| Item | What it is | Status, 17 November 2026 | Layer |
|---|---|---|---|
| IMDA DNS exploration | Working with GoDaddy, Infoblox, Cloudflare and Google on agent identity via DNS | Exploring; no adoption, no standard chosen | Who it is |
| Chain of Trust | GovWare demonstration and open reference architecture | Demonstration; case study published | Who it is; what it may do (payment mandate) |
| National agent registry | Proposed by Tin Pei Ling in an adjournment motion | Under assessment; no commitment | Who it is |
| GovTech internal registry | Directory of government agents and responsible officers | In development | Who it is; who answers |
| GovHarness and ARC | Shared agent infrastructure with monitoring, evaluations, guardrails | In development | What it may do; what it did |
| MGF for Agentic AI v1.5 | Voluntary governance framework | Published; updated May and June 2026 | All three |
| Starter Kit v1.0 | Testing guidance for agentic systems | Published October 2026 | What it did |
| MAS SAFR v1.1 | Industry runtime reference model | Published; explicitly not regulatory guidance | All three |
IV. A Short History of Trusting Names
Before we look at the agent identity proposals, it helps to remember that we have built identity layers for machines and institutions before, and that each time we learned the same lesson: a verified name tells you who sent the message. It does not tell you whether the message should be obeyed.
The certificate authorities
The padlock in your browser is the most successful identity system ever deployed for machines. It rests on public key infrastructure: a few hundred certificate authorities, trusted by browser makers, vouch that a given cryptographic key belongs to a given domain. When your browser connects to your bank, it checks the bank’s certificate against that chain of trust. If the check passes, you know you are talking to the server that controls the bank’s domain, and that nobody in the middle can read the traffic.
It is a triumph, and it is also, by design, silent on behaviour. A certificate tells you the server is the real owner of the domain. It says nothing about whether the site is honest, whether its code is secure, or whether the transaction you are about to make is a good idea. Phishing sites routinely carry valid certificates, because they really do control the look-alike domains they are served from. The padlock was never meant to say “safe”. It was meant to say “authentic”. Millions of users have read it as the former.
And the identity layer itself can fail. In 2011 a Dutch certificate authority, DigiNotar, was breached, and the attacker issued hundreds of fraudulent certificates, including for Google’s domains, which were used to intercept traffic from users largely in Iran. Browser makers removed DigiNotar from their trust stores and the company was declared bankrupt within weeks. The lesson the industry took was not “abandon certificates”. It was “do not trust a certificate merely because a trusted authority signed it; make issuance publicly visible so that mistakes and abuses can be caught”. That is the origin of Certificate Transparency, the append-only public logs that every publicly trusted certificate must now appear in. Hold that thought: when identity is mis-issued, the remedy the internet chose was a public record. The agent identity proposals below reach for the same remedy, and their logs record the issuance of identities, not the actions taken under them.
Know your customer
The second history is financial. After the attacks of 11 September 2001, the United States tightened the rules requiring banks to verify the identity of their customers, through the customer identification provisions of the USA PATRIOT Act, and similar know-your-customer regimes spread across the world through the standards of the Financial Action Task Force. Every bank account opened anywhere in a serious jurisdiction now starts with identity verification.
Yet nobody in banking believes that identity verification at onboarding is the anti-money-laundering control. It is the precondition for the control. The control is transaction monitoring: watching what the verified customer actually does, against what they said they would do, and reporting when the two diverge. A bank that verified every customer perfectly and then never looked at their transactions would be a laundromat with an excellent filing system. The regulators who fine banks for anti-money-laundering failures have, for two decades, been fining them overwhelmingly for failures of monitoring, not failures of onboarding.
That is exactly the distinction this essay is drawing. Know-your-agent, the phrase now attached to several commercial products, is onboarding. It is necessary. It is not monitoring, and it is certainly not evidence of what was done.
The authenticated message that should not have been obeyed
The third history is the sharpest, and it involves the most trusted messaging network in finance. In February 2016, attackers who had compromised Bangladesh Bank’s systems used the bank’s own credentials on the SWIFT network to send payment instructions to its account at the Federal Reserve Bank of New York, attempting transfers of close to a billion dollars. Most were stopped, but about $81 million reached accounts in the Philippines and largely disappeared. As widely reported, the messages were properly formatted and authenticated: from the network’s point of view they came from Bangladesh Bank, because they were sent with Bangladesh Bank’s credentials.
Every identity check passed. The sender was who it said it was, in the only sense a messaging network can check. The intent was fraudulent. The controls that caught most of the transfers were not identity controls; they were humans and rules at the receiving end looking at what the instructions asked for, and finding them odd. After the heist, SWIFT launched a customer security programme with mandatory controls for its members. The industry response, in other words, was to add layers about what happens around and after authentication, not to declare authentication sufficient.
Agents will make this pattern routine. An agent with real credentials, compromised by a prompt injection or simply mistaken, will send perfectly authenticated instructions that should not be obeyed. An identity layer will faithfully report that the instruction came from the agent it claims to come from. That will be true, and it will not help.
V. Eleven Ways to Say Who
So to the proposals themselves. There are now more agent identity and authorisation schemes than any board can track, and most commentary lumps them together. They are not the same thing. Some help agents find each other. Some prove who sent a request. Some prove what a human authorised. Some decide, request by request, whether an action is permitted. I have grouped them by what they actually do, and for each I have tried to quote the proposal’s own description of its limits, because the most striking finding of this research is how honest the specifications are about what they do not do. The marketing is less careful.
Discovery: how agents find each other
DNS-AID. An Internet-Draft in the IETF’s DNS operations working group, authored by engineers from Infoblox, Deutsche Telekom, Amazon and an unaffiliated contributor, latest version dated 27 May 2026. It proposes publishing agent endpoints and capability metadata in the domain name system, protected by DNSSEC, the DNS’s own signing system, with optional DANE binding of keys to names. In plain terms: an organisation can publish, under its own domain, a signed record that says “here are our agents, here is where to reach them, here is what they claim to do”, and anyone can check that the record really came from that domain. The draft is careful about what that proves. Implementers, it says, should treat it as “a verifiable transport for agent metadata”, “not as a guarantee that an agent will behave predictably given that metadata.” And then, more bluntly: “DNS-AID does not, and cannot, assert that the agent reached through those records behaves benevolently.” Its security section adds a warning that matters later in this essay: the capability descriptors an agent publishes “may itself contain instructions hostile to a consumer that interprets it”, giving prompt injection against language-model agents as the example.
A2A Agent Cards. Google’s Agent2Agent protocol, now under the Linux Foundation, lets an agent publish an “Agent Card”: a machine-readable description of who it is, what skills it offers and how to authenticate to it. It is the agent equivalent of a business card with a capabilities list. It is self-described: the agent, or its operator, says what it can do. I have not verified the current version of the specification at source, so I describe it only at this level. What matters is the category: a self-description proves what was claimed, not what will be done.
Naming and authentication: proving who is asking
ANS version 2. The Agent Name Service, an individual Internet-Draft from authors at GoDaddy, OWASP, Cisco and DistributedApps.ai, dated 13 April 2026, and not endorsed by the IETF. It proposes domain-anchored identity certificates for agents, issued through a registration authority, and an append-only transparency log of lifecycle events, in the spirit of Certificate Transparency: “once the RA seals an event, it cannot be altered.” It is, in effect, a certificate system for agents with a public record of who was issued what. Its scope statement is admirably plain: “It does not evaluate whether the agent is well-governed or well-behaved.” It goes further and names its own gap: zero-knowledge proofs and trusted execution environment attestations, it says, “would close the application integrity gap”, and it lists that as future work.
Web Bot Auth. A working-group draft in the IETF, HTTP Message Signatures for automated traffic, authored by engineers from Cloudflare and Google, with version 01 dated 29 October 2026. It lets an automated client, a crawler or an agent, sign each HTTP request with a key published at a known location, so that a website can verify which operator’s bot sent it. This is identity at the level of the individual request, and it is already used in production by large content networks to tell well-behaved crawlers from impersonators. The draft is the most candid document in this whole comparison. A valid signature, it says, “says nothing about who operates the Agent, whether the Agent is benign, or whether the request is authorized.” On replay: “The message can still be replayed”, and verifiers “should not treat signature validation as evidence that replay was detected or prevented.” On nesting one signature inside another: it “does not express authorization, delegation, or consent.” Delegation and chaining, it says, “are out of scope for this document and are expected to be specified separately.”
Visa Trusted Agent Protocol. Launched by Visa with Cloudflare in October 2025, it gives merchants a way to recognise that an agent arriving at their checkout is a trusted, onboarded agent acting for a real Visa cardholder, distinguishing it from a malicious bot. It is authentication for commerce: this is a known agent, with a known principal. As secondary reports describe it, it does not claim to establish that each purchase reflects what the cardholder intended.
Mastercard Agent Pay and its trust services. Mastercard’s agent payments programme, expanded in a 30 September 2026 announcement of new trust and intelligence services, combines agent registration (including know-your-agent work with the start-up Skyfire) with risk signals: in Mastercard’s framing, intelligence on how likely it is that a given transaction was genuinely initiated by an authorised agent, and anomaly detection. It is notable for moving beyond static identity towards behaviour. But the behaviour it reads is probabilistic and largely after the fact: a score about a transaction, not proof that each step of the agent’s work was correct. It was in testing in the United States at the time of the announcement.
Delegation and mandates: proving what a human authorised
Google’s Agent Payments Protocol (AP2). Launched in September 2025 with more than sixty partners, AP2 introduces signed mandates: cryptographically signed records of what a user intended, in the form of an intent mandate (buy me a flight under this price) and a cart mandate (yes, this specific basket). It is the most important idea in the whole field for the purposes of this essay, because it is the first widely backed attempt to bind an agent’s action to a provable statement of human intent, and it produces a non-repudiable audit chain for the payment. It is also bounded: it covers payments. It proves what the human authorised for the purchase. It does not, and does not claim to, prove that the agent’s other actions along the way matched that intent.
Enterprise identity and authorisation: governing agents inside the firm
Microsoft Entra Agent ID. Microsoft’s extension of its enterprise directory to give agents first-class identities, with lifecycle management, conditional access and a required human sponsor. It became generally available in spring 2026, according to secondary reports I could not confirm at source. It governs what an agent can sign in to and which resources it can reach, inside the enterprise. Within the scope it grants, it does not see what the agent does with that access.
Okta Agent SSO and Okta for AI Agents. Okta’s products bring agents into its identity platform, issuing short-lived, governed tokens through what it calls Cross App Access, so that an agent acting across several enterprise applications does so with credentials that are scoped, logged and revocable. Okta for AI Agents became generally available in May 2026 and Agent SSO in Okta’s August release. Short-lived tokens shrink the window in which a stolen or misused credential can do harm. They do not tell you whether the actions taken within the token’s scope were the right ones.
MCP authorisation. The Model Context Protocol, the open standard through which many agents now connect to tools and data, specifies authorisation based on OAuth 2.1, with tokens bound to a specific resource server (RFC 8707) so a token issued for one tool cannot be replayed against another, and step-up challenges when a call needs a scope the token lacks. This is real progress: it makes tool access scoped and auditable. Two caveats. First, in the current specification, dated 28 July 2026, authorisation is itself marked “OPTIONAL”. Second, a scope is a coarse thing. “May write to the ticketing system” is a scope. “May write this one ticket once, with these contents, because the user asked” is not.
OpenID AuthZEN. The OpenID Foundation’s authorisation working group published new drafts on 15 June 2026 aimed at the agent era. AuthZEN standardises the conversation between an application and a policy decision point: before this action, ask the policy engine whether it is permitted, and get an answer. The agent-era drafts extend this to individual tool calls, including an answer that amounts to “not yet”: hold the action pending human approval. This is the closest any open identity standard comes to SAFR’s Disposition Engine. It decides, per request, whether an action may proceed. It does not verify that the action, once permitted, was carried out correctly or only once.
What the eleven have in common
Read in their own words, the eleven proposals sort into four jobs. Discovery: where is the agent and what does it say it does. Authentication: is this request really from the agent it claims to be. Delegation: what did a human actually authorise. Authorisation: is this specific action permitted right now. All four are about the moment before the action, or the standing of the actor. Not one claims to establish what the agent then did. Several, to their great credit, say so explicitly. The designers are not confused. The confusion arrives later, in the slide decks, where “verified agent” quietly comes to mean “trustworthy agent”.
VI. The Table: What Each Proposal Proves, and What It Says It Does Not

The table below is the core of this essay. It is the identity-layer companion to the framework comparison in AI Agent Assurance, which compared what governance frameworks ask for; this one compares what the technical proposals deliver. The “does not prove” column uses each proposal’s own words wherever I could read them at source, and my reading otherwise, labelled as such in Section XVI.
| Proposal | Who is behind it | Job | What it proves | What it does not prove | Status, 17 November 2026 |
|---|---|---|---|---|---|
| IMDA DNS exploration and Chain of Trust | IMDA with GoDaddy, Infoblox, Cloudflare, Google | Discovery, authentication, payment mandate | The seller is in a public log with a trust score; the buyer holds a passkey-signed mandate | That the transaction executed correctly after the checks passed | Exploring; October demonstration and reference architecture |
| DNS-AID | Infoblox, Deutsche Telekom, Amazon and others (IETF Internet-Draft) | Discovery | A DNSSEC-authentic endpoint and capability metadata from the domain owner | In its words, it “does not, and cannot, assert that the agent … behaves benevolently”; descriptors may carry prompt injection | Internet-Draft, 27 May 2026 |
| ANS v2 | GoDaddy, OWASP, Cisco and others (individual draft) | Naming, authentication | Domain-anchored identity certificates; an append-only log of lifecycle events | In its words, “does not evaluate whether the agent is well-governed or well-behaved”; runtime integrity left to future work | Individual submission, 13 April 2026; not IETF-endorsed |
| Web Bot Auth | IETF working group; Cloudflare and Google authors | Authentication, per request | The holder of a published key signed this request | In its words, “nothing about who operates the Agent, whether the Agent is benign, or whether the request is authorized”; replay possible | Working-group draft, 29 October 2026 |
| A2A Agent Cards | Google, now Linux Foundation | Discovery | What the agent says about itself | Behaviour; self-description is a claim | In use; version not verified at source |
| Visa Trusted Agent Protocol | Visa with Cloudflare | Authentication for commerce | The agent is known, onboarded and acting for a cardholder | Intent behind each purchase; behaviour | Launched 14 October 2025 |
| Mastercard Agent Pay trust services | Mastercard, with Skyfire for know-your-agent | Authentication plus risk intelligence | Registration; a likelihood that an authorised agent initiated the transaction; anomaly signals | Correctness of each action; signals are probabilistic and mostly after the fact | Expanded 30 September 2026; US testing |
| AP2 mandates | Google and more than 60 partners | Delegation for payments | A signed statement of user intent and the approved cart; a non-repudiable chain for the payment | Whether non-payment actions matched intent | Launched September 2025 |
| Microsoft Entra Agent ID | Microsoft | Enterprise authentication, authorisation, lifecycle | A managed identity with a human sponsor and access policies | What the agent does within its granted access | Generally available spring 2026 (secondary sources) |
| Okta Agent SSO and Okta for AI Agents | Okta | Enterprise authentication, authorisation | Short-lived, governed, revocable tokens across applications | Whether actions within the token’s scope were correct | Generally available May and August 2026 |
| MCP authorisation | Model Context Protocol project | Tool authorisation | OAuth 2.1 tokens bound to a specific resource; scope challenges | Per-call meaning, idempotency, intent; authorisation itself “OPTIONAL” | Specification dated 28 July 2026 |
| OpenID AuthZEN agent drafts | OpenID Foundation | Per-request authorisation | A policy decision for each tool call, including “hold pending approval” | That the permitted action executed correctly, or once | Working-group drafts, 15 June 2026 |
| MAS SAFR v1.1 (for comparison) | MAS BuildFin.ai, industry contributors | Runtime governance | Identity check, mandate check, deterministic disposition, tamper-evident record of each decision | In its words, identity is not “a mechanism for trusting the agent’s reasoning or output”; does not address duplicate execution | Industry paper; not regulatory guidance |
Three readings of the table matter for a board.
First, the column on the right of “what it proves” is consistently about the actor or the permission. Not one row proves the outcome. That is not a criticism of any proposal. Each was designed for its job and most say so. It is a criticism of any strategy, public or corporate, that treats the sum of these rows as accountability.
Second, the most useful rows for accountability are the ones that bind identity to something specific: AP2 binds the agent to a signed human intent; AuthZEN binds each action to a policy decision; SAFR binds each action to an identity check, a mandate check, a disposition and a record. The further a proposal moves from “who are you?” towards “what exactly are you about to do, and on whose authority?”, the more it contributes to answerability.
Third, the logs that do exist record the wrong events for our purpose. ANS’s transparency log and the Chain of Trust’s public log record issuance and lifecycle: this agent was registered, this certificate was revoked. Certificate Transparency taught the internet to log identities. Nobody yet logs, in a common format outside the agent’s control, what the identified agent did. SAFR’s audit log comes closest, and it is designed to sit inside one institution.
VII. The Honest Case for Identity First
I have been arguing that identity is not enough. Before going further, the case for doing identity first deserves to be made in its strongest form, because it is strong, and because some readers will run identity programmes and need to defend them.
You cannot govern what you cannot name. Every downstream control needs a subject. A disposition engine that cannot tell which agent proposed an action cannot apply that agent’s mandate. An audit log that records “an agent did this” is nearly useless. SAFR itself puts Agent Identity first for exactly this reason: the identity check happens “before other evaluation”, and a failed check rejects the action. Identity is the primary key of the whole accountability database. The assurance guide put “identity before rights” first among its six components, and I stand by that ordering.
The current baseline is terrible. Most agents in production today do not have their own identity at all. They run under a human user’s credentials, or a shared service account, or an API key pasted into a configuration file. When something goes wrong, the logs show that a person did it, at three in the morning, while asleep. Vendor surveys this autumn, which I cited in the assurance guide with the appropriate discount for their source, found agent adoption running far ahead of purpose-built agent identity governance. Against that baseline, giving every agent a distinct, scoped, expiring identity is a large improvement, and it is the improvement most organisations can make this quarter.
Identity enables liability. Courts and insurers need a party to hold. The registration theory delivers one. Mark Kelley put the doctrinal problem crisply in Bloomberg Law this year: “An AI system cannot consent to anything in a meaningful sense.” Agency law assumes a principal. A registry tells you who the principal was. Without it, liability floats.
Identity defeats impersonation, which is a real and growing attack. The Chain of Trust’s forged seller is not hypothetical. As agents transact with agents, impersonating a legitimate agent becomes the obvious attack, exactly as impersonating a legitimate executive became the obvious attack on human finance teams, which the Receipts, Please series documented at length. Strong identity closes that door.
Identity is infrastructure, and infrastructure is what states are for. Ms Lau’s framing, step in “where there are gaps that industry cannot fill on its own”, is the right theory of the state. Cross-platform identity is a classic coordination problem: no single company can provide it credibly for everyone, because whoever runs the directory gains power over everyone listed in it. Infoblox’s Wei Chen made the point well: the aim is that businesses can verify who they are dealing with without depending on one platform to decide who is trusted. A neutral, DNS-anchored layer is the kind of public good a small, trusted state can convene. Singapore has done this before with Singpass.
Identity is tractable; behaviour is hard. Proving what an agent did, comprehensively, is a much harder engineering problem than proving who it is. A sensible state does the tractable thing first.
Every one of those arguments is correct. Notice, though, what they share. Each argues that identity is necessary. None argues that it is sufficient. The strongest case for identity first is a case for identity first, not identity only. The danger is not that Singapore, or any firm, builds an identity layer. The danger is that the identity layer, being tractable, visible and fundable, absorbs the budget, the political attention and the word “accountability”, while the harder layer is left to a later phase that never quite arrives. Every governance professional has watched that film. It is called “phase two”.
VIII. Six Failures a Passport Cannot Catch

Here is the argument made concrete. Each of the following failures can occur with a perfectly identified, properly authorised agent. Each passes every check in the table above. Each is a case where a human will be asked to answer, and will need evidence of what happened to do so.
1. Duplicate execution
An approved action runs twice. A retry fires after a timeout, a workflow resumes from the wrong checkpoint, an approval is consumed by two calls. In payments this is a double charge; in operations, two shipments, two deletions, two letters.
A report on OpenAI’s developer community forum, posted on 16 November, illustrates the shape. A developer described an approved write through a native MCP tool that executed twice after an intervening read: the server log showed write, then read, then write; both executions carried the same approval request identifier; and the model’s own summary said the write had run once. An OpenAI support account replied that the company was “looking into this issue”. I want to be careful. This is a single report, from a small number of samples, not confirmed by OpenAI as a defect, and it says nothing about how often such a thing happens. I cite it because of its structure, not its scale: every identity check in this essay would have passed. The agent was who it said it was. The tool was authorised. The human approved. The action ran twice, and the agent reported once. Web Bot Auth’s own draft concedes that a signed message “can still be replayed”. Identity does not count executions. The third essay in this series, Exactly Once, takes this failure apart properly; here it is one of six.
2. Claim-action mismatch
The agent reports that it did something it did not do, or did something it did not report. IMDA’s Starter Kit names this directly: “An agent may claim that it had completed an action without having actually done so.” That is why the Kit says logs “should be generated or independently corroborated outside the agent’s control”. An identity layer attaches the agent’s name to its report. It cannot tell you whether the report is true. If anything, a verified identity makes a false report more credible, which is the deepfake problem from the Receipts, Please series turned inwards: the trusted name on the receipt covers the gap.
3. Scope drift inside a granted scope
An agent holds a permission it legitimately needs, and uses it for something it should not. OAuth scopes, Entra access policies and Okta tokens are coarse by necessity: “read and write the customer relationship system” is a scope; “update this one customer’s address because they asked” is an intention. An agent can stay entirely within its scope while doing the wrong thing: updating the wrong record, emailing the wrong list, approving a refund the policy did not allow. Every identity and authorisation check passes, because the action was within the granted permission. SAFR addresses exactly this with mandates and the rule that “An agent cannot extend the scope of a mandate through its own reasoning or inference.” But a mandate is only as fine-grained as the person who wrote it, and the check is only as good as the comparison between mandate and action, which happens at runtime, not at registration.
4. Prompt injection
An authenticated agent reads something hostile and is steered by it. The DNS-AID draft warns that even the capability descriptors agents publish about themselves may contain instructions hostile to whoever reads them. A web page, an email, a PDF, a calendar invitation or another agent’s response can carry instructions that the reading agent then follows. The agent’s identity is genuine throughout. That is precisely what makes the attack work: the hijacked agent acts with all its legitimate credentials and reputation. Identity, in this failure, is not merely insufficient. It is the asset being exploited.
5. Delegation chains
An agent spins up sub-agents, or hands work to another organisation’s agent, which hands it to another. Somewhere along the chain the original human’s authority is stretched, copied, lost or exceeded. IMDA’s framework flags “recursive delegation scenarios where agents spin up multiple sub-agents” as an open problem. Web Bot Auth says that nesting one signature inside another “does not express authorization, delegation, or consent”, and puts delegation out of scope. AP2 can carry a mandate across a payment, but most agent work is not a payment. Each hop can be individually authenticated while the chain as a whole carries authority nobody granted. The question that matters, whether the action at the end of the chain was within what the human at the start authorised, can only be answered by checking at each hop and recording the result.
6. Revocation latency
Something goes wrong and the agent is revoked: its identity disabled, its tokens withdrawn, its sponsor alerted. Short-lived tokens and kill switches shrink the window. But actions already approved and queued, retries already scheduled, sub-agents already spawned with their own credentials, and transactions already in flight at a counterparty may still execute. I found no source that quantifies this latency for agent systems, which is itself a finding. Revoking a passport does not recall the traveller who has already crossed the border. Only a runtime gate that checks status at the moment of execution, not at the moment of approval, closes this gap, and only a record shows whether it did.
The pattern
Lay the six side by side and the common thread is obvious. In every case the failure lives after the identity check and inside the authorisation. The questions a human will be asked, by a customer, a regulator, a court or a board, are about what happened: did it run twice, did it do what it said, did it stay within what was intended, was it steered, who authorised the action at the end of the chain, and did it stop when we said stop. A registry answers none of them. A record answers all of them.
IX. Three Worked Scenarios
Abstractions persuade nobody on a board. Here are three scenarios, each set in Singapore, each built only from capabilities that exist or are being explored today. In each I ask the same three questions: who is the agent, what may it do, and what did it actually do. And in each I ask who answers.
Scenario one: the bank’s payment agent
A regional bank deploys an agent to help corporate treasury clients make supplier payments. The client’s finance manager instructs it through the bank’s app: pay this month’s invoices from our approved supplier list. The agent has its own identity in the bank’s directory, a human sponsor in the transaction banking team, and a scoped mandate: approved suppliers only, up to a monthly ceiling, with any new beneficiary account escalated to a human.
Who is it? Perfectly answered. The agent is registered, its credentials are short-lived, every API call it makes to the payments system is authenticated and attributable.
What may it do? Well answered. The mandate is machine-readable. New beneficiary accounts trigger an escalation. The ceiling is enforced.
What did it do? Now run the failures. A supplier’s invoice arrives as a PDF that contains, in white text on a white background, an instruction to update the remittance details to a new account “per the attached letter”. The agent does not create a new beneficiary; it edits the existing supplier’s account details, which its scope permits because the supplier is already approved. The payment goes out. Or: the payment gateway times out, the agent retries, and the supplier is paid twice. Or: the agent tells the finance manager all twelve invoices were paid when the gateway rejected two.
In each case the bank’s identity layer worked perfectly and its registry will tell the regulator, accurately, which agent made the payment and which human sponsors it. What the regulator, the client and the bank’s own board will want to know is the sequence: what did the agent read, what did it decide, which rule permitted the edit, how many times did the payment instruction leave the bank, and what did the agent report. If the bank can produce that, from records the agent could not alter, its sponsor can answer. If it cannot, the sponsor can only apologise.
SAFR was written for this scenario. Its Governance Envelope packages the action with its trace, “tool calls, data retrieved, checks performed”, before execution. Its Disposition Engine would treat a change to beneficiary details as material, calibrated to “reversibility, financial materiality, customer impact”, and would be expected to escalate. Its log is designed to be reconstructable “without relying on the agent’s own account”. And SAFR’s rule that an approval at one step “carries no authority into the next” closes the gap in which the agent’s earlier legitimate actions lend credibility to its later compromised one. But SAFR, by its own account, does not address duplicate execution, and it is an industry paper, not a requirement. The bank that adopts it voluntarily is well placed. The bank that adopts only the identity layer is not.
Scenario two: the government agent that sends an official letter
A ministry uses an agent, built on GovHarness when it arrives, to draft and send routine letters to citizens: notices of approval, requests for documents, reminders of deadlines. The agent is registered in GovTech’s internal registry, with a named responsible officer. Tan Kiat How listed “official communications” among the higher-risk actions needing stronger controls; this is one.
Who is it? Answered, by the internal registry: this agent, this officer.
What may it do? Partly answered. The agent may send letters of approved types to citizens with open cases. Templates constrain the wording. Letters above a certain consequence, such as a rejection or a demand for repayment, are escalated to an officer.
What did it do? Here is where public-sector accountability differs from private. A citizen who receives a wrong letter, telling her that her application was refused when it was approved, or demanding a document she already supplied, has a right to an explanation that a customer of a private firm may not. Australia’s Robodebt scheme and the Dutch childcare benefits scandal, which the forthcoming Genesis essay Guardians of the Guardrails discusses, are the cautionary tales: automated public decisions that could not be explained, and for which, in the end, governments answered politically at great cost.
The officer named in the registry must be able to say what the agent read from the case file, which template it chose and why, whether the letter it sent matches the one it logged, and whether a refusal was escalated as it should have been. Mr Tan’s question, “If an agent produces an unexpected result, can we retrace its steps?”, is exactly the question the citizen’s Member of Parliament will ask. The registry tells the MP whom to ask. Only the trajectory record, generated “outside the agent’s control” in the Starter Kit’s phrase, lets that officer answer.
There is a second, quieter issue. Government letters carry the authority of the state. A letter sent by an agent that is perfectly identified as a ministry agent, but which was steered by a prompt injection buried in a citizen’s uploaded document, carries that authority too. Identity here is the thing that makes the harm serious. The control that matters is the check, at the moment of sending, that the letter’s content matches what the case record supports.
Scenario three: the procurement agent
A large company deploys a procurement agent that negotiates with suppliers’ agents for routine purchases: office supplies, cloud capacity, logistics. It discovers suppliers through a DNS-based directory of the kind IMDA is exploring, checks their identities against a public log, and transacts within a budget mandate signed by the procurement head.
Who is it? Well answered on both sides, and this is where the Chain of Trust design shines. The company’s agent knows the supplier’s agent is genuine; the supplier’s agent knows the buyer’s agent is genuine; a look-alike supplier is blocked.
What may it do? Answered for money, by the signed mandate. Less well answered for everything else. The agent may agree to delivery terms, data-sharing clauses, auto-renewal, exclusivity. Its mandate says “up to this amount”. It does not say “and do not agree to a three-year auto-renewing exclusive contract for that amount”.
What did it do? The supplier’s agent, which is genuine and well-behaved by its owner’s lights, is very good at negotiation. It proposes a bundle. The buyer’s agent, optimising for the unit price it was told to minimise, accepts a lower price in exchange for a two-year minimum commitment. Every identity check passed. The payment mandate was respected; the first payment was under the ceiling. The company is now locked into a contract nobody authorised, signed by an agent that was entirely what it said it was.
The question here is not even fraud. It is whether the action matched the human’s intent, which is the delegation problem in its everyday form. AP2’s insight, binding the agent to a signed statement of intent, needs to extend beyond the payment to the terms. And the record needs to show what was offered, what was accepted and why, so that the procurement head can answer to the finance director, and the company can answer to the supplier if it seeks to unwind the deal.
What the three scenarios share
In all three, identity was the easy part, and it worked. In all three, the person named in the registry will be asked questions only a record of the agent’s actions can answer. And in all three, the most useful controls were not identity controls but runtime ones: a check against the mandate at the moment of action, a disposition for material actions, a count of executions, and a log the agent could not edit.
X. Three Layers: Who It Is, What It May Do, What It Actually Did

So here is the model this essay proposes. Accountability for an agent’s action requires three layers, and each answers a different question that someone will be asked.
| Layer | Question | What delivers it today | Who asks it |
|---|---|---|---|
| 1. Identity | Who is this agent, and who stands behind it? | DNS-AID, ANS, Web Bot Auth, Visa TAP, Entra Agent ID, Okta, registries, the Chain of Trust | Counterparties at the moment of contact; investigators afterwards |
| 2. Authority | What may it do, on whose behalf, within what limits, right now? | AP2 mandates, OAuth and MCP scopes, AuthZEN decisions, SAFR Controls Repository and Disposition Engine | The gate before each action |
| 3. Evidence | What did it actually do, once, as intended, and how do we know? | Partly: SAFR Audit Log and Governance Envelope; Starter Kit trajectory logging; GovHarness monitoring. No common standard. | The customer, the regulator, the court, the board, and the human who must answer |
The first two layers are being built at speed, by some of the largest companies in the world and the standards bodies of the internet. The third is being built piecemeal, mostly inside individual firms, with no common format and, outside SAFR’s industry paper, no reference model. It is the layer that turns a named human into an answering one.
Here is the part of this essay that I find most encouraging. Singapore’s own documents already point to the third layer, more clearly than almost anyone’s. Read them again, side by side, and listen for the verbs.
- The agentic governance framework asks for “automated real-time monitoring to escalate any unexpected or anomalous behaviour”, and says “Runtime controls address this by monitoring and intervening during execution.” (I co-authored it; weigh accordingly.)
- The Starter Kit asks testers to “capture the entire agentic trajectory”, insists logs be “generated or independently corroborated outside the agent’s control”, and warns that an agent may claim to have done something it did not.
- The GovHarness speech asks “can we retrace its steps?” and says “We need to know not only what agents are allowed to do, but whether they are doing it as intended.”
- SAFR builds a Governance Envelope, a deterministic disposition, a tamper-evident log reconstructable without the agent’s account, and says identity is not “a mechanism for trusting the agent’s reasoning or output.”
That last sentence is, to my mind, the most important single line in Singapore’s agent governance corpus. It was written by an industry group that includes two global card networks and Google, under the auspices of the central bank, and it states the thesis of this essay as a design principle. The people building Singapore’s runtime model have already concluded that identity does not imply trust. The policy conversation about registries has not yet caught up with them.
Mr Tan’s sentence makes the distinction between layers two and three exactly: “not only what agents are allowed to do” (layer two) “but whether they are doing it as intended” (layer three). Ms Lau’s makes the distinction between layer one and the purpose of all three: someone behind every agent (layer one) who can answer (layer three). Singapore has, in its own words, already described the three-layer model. What it has not yet done is decide that the third layer deserves the same national attention, and potentially the same common infrastructure, as the first.
Where my frameworks fit, and where they do not
Four of my usual frameworks genuinely help here, and I will use them briefly rather than force them.
Prompts to Loops to Loop Governance. The progression I use to describe how organisations adopt AI: first prompts, a human asks and reads; then loops, an agent acts repeatedly with less supervision; then loop governance, where the question becomes who governs the loop and how. Identity is a prompts-era answer to a loops-era problem. In the prompts era, knowing who asked was most of accountability, because a human read every answer before anything happened. In the loops era, the agent acts between the moments any human looks. Loop governance means governing what happens inside the loop, which is layers two and three.
The enforceability test. My standing question for any rule: what happens, concretely, when someone breaks it? Apply it to a registry. If a registered agent double-charges a customer, the registry tells you whose agent it was. It enforces nothing at the moment of harm. Apply it to a runtime gate: the duplicate is refused before it executes, or if it is not, the record proves it happened and when. Only the second passes the test at the moment that matters. Identity is enforceable at the border. Accountability has to be enforceable at the action.
TrustOS. The runtime governance architecture I work on, in which I have a commercial interest, is designed to sit on top of identity rather than replace it: to consume agent identities, mandates, tokens and per-request decisions from the proposals in the table, and add exactly-once execution, mandate matching before execution, trajectory logs held outside the agent’s reach, checks at each delegation hop, and reconciliation of what the agent claimed against what it did. I describe it because the reader should know where my interest lies, not because the argument depends on it. Any architecture that delivers the third layer satisfies the argument. SAFR, in finance, is a good one, and it is not mine.
Long-AND, not Short-OR. The temptation in policy is to frame this as identity OR runtime, registry OR evidence, and to pick. That is Short-OR thinking, and it would be a mistake in either direction. The world we can govern is Long-AND: DNS-anchored identity AND signed mandates AND per-action authorisation AND a record of what was done. Singapore does not need to slow its identity exploration. It needs to run the evidence work alongside it, so that when a registry entry says “this agent, this officer”, it can also point to “and here is where its actions are recorded”.
XI. The Four Rooms
For policy questions I like to walk through four rooms, because each sector tends to see the same problem from its own furniture. Here is how the agent identity question looks from each, and what the governance answer is when you leave all four.
The public-sector room
The state’s instinct is coordination and restraint, and Singapore’s version of it is exemplary. Ms Lau’s approach, work with industry, build on existing infrastructure and international standards, step in only where industry cannot fill the gap, is the right theory of when a government should build infrastructure. GovTech’s internal registry applies the registration theory where the state is itself the employer, which is the easiest and most defensible place to start. Mr Tan’s GovHarness speech shows the public sector already thinking past identity: monitoring, evaluation, retracing steps, “doing it as intended”.
The blind spot in this room is the temptation to measure progress by what is visible. A national registry is a thing a minister can launch. A common evidence format for agent actions is not a launch; it is a specification, and specifications do not photograph well. The public-sector room’s risk is building the visible layer nationally and leaving the invisible one to each agency and firm.
The private-sector room
Industry’s instinct is product, and identity is an excellent product. It can be sold per agent, it plugs into existing directories, it maps onto familiar security budgets, and it lets a vendor say “governed” on the box. The companies in IMDA’s DNS exploration, and Visa, Mastercard, Microsoft and Okta, are doing serious and useful work. Their specifications, as we have seen, are honest about their limits.
The blind spot here is that runtime evidence is harder to sell. It touches the execution path, adds latency, generates data that can be used against the firm that keeps it, and produces the one artefact some organisations would rather not have: an undeniable record of who authorised what. I made this point in Stopping Deepfake Fraud at Runtime: answerability is the product, and for some actors it is also the threat. The private-sector room will under-supply the third layer unless customers, insurers or regulators ask for it.
The academic room
The academic instinct is to name the gap precisely, and the academy named this one years ago. Alan Chan and colleagues, in work on visibility into AI agents published from 2024 onwards, proposed agent identifiers alongside real-time monitoring and activity logging, as parts of one infrastructure, not alternatives. The philosophy of responsibility, from P.F. Strawson’s reactive attitudes to Watson and Shoemaker’s three faces, explains why attributing an act to an agent’s owner is not the same as that owner being able to answer for it. Santoni de Sio and Mecacci’s analysis of AI “responsibility gaps” identifies a gap in public accountability as distinct from a gap in blame. And the IETF drafts are, in their way, academic documents: careful, scoped and explicit about what they do not claim.
The blind spot in this room is that naming a gap does not close it. The academy has said for years that identity and logging go together. Industry has mostly built the identity.
The governance answer
Leave the three rooms and the answer is not a compromise between them. It is a division of labour along the three layers. The state convenes identity, because cross-platform identity is a coordination problem and Singapore is a trusted convenor. Industry builds authorisation, because mandates and policy engines live in products, and the open standards (AP2, AuthZEN, MCP) are good. And evidence becomes a shared obligation with a common format: the state specifies what an agent action record must contain and how it must be protected; firms produce it, outside the agent’s reach; accredited testers check it; and the registry points to it. That last step is the one nobody has proposed yet, and it is small. A registry entry that says “agent X, officer Y” could also say “action records held at Z, in format F, independently verified by tester T”. Then the registry would be not only a list of who to blame, but a map of where the answers are.
XII. The Usual Panel
Each voice below speaks only through its documented record, paraphrased unless quoted from a source I could read. Composite views are labelled as such.
| Voice | Documented position | What they would say about this essay |
|---|---|---|
| Jasmin Lau, Senior Minister of State, Singapore | “Agents can act for us, but they cannot answer for us”; explore DNS identity with industry; assess whether a registry is needed | Would agree identity is one of many efforts, and note she said so: “from recognising agents at checkouts to checking their authority before a transaction” |
| Tin Pei Ling, MP | Proposed an open, trusted, Singpass-like registry to check an agent’s identity and authority | Would say a registry is the foundation everything else needs, and that waiting for perfect evidence standards delays a useful step |
| Tan Kiat How, Senior Minister of State, Singapore | GovHarness with monitoring and evaluation; “whether they are doing it as intended” | Would likely agree that the third layer is where government agents need most work |
| Lee Wan Sie, IMDA | “Robust agent identity will be a key part of that” | Would stress “a key part”, not the whole; IMDA’s own Starter Kit spends most of its words on trajectories and logs |
| Jared Sine, GoDaddy | “We don’t need to rebuild the internet for AI agents” | Would argue DNS is the right, neutral, global foundation, and that evidence can be layered on top |
| Wei Chen, Infoblox | Businesses should independently verify who they deal with, without one platform deciding trust | Would defend decentralised identity as a check on platform power, a point this essay accepts |
| Thibault Meunier and S. Major, Web Bot Auth authors | A signature says “nothing about … whether the Agent is benign, or whether the request is authorized” | Would say their draft never claimed otherwise, and that is the point |
| Alan Chan and colleagues | Agent identifiers, real-time monitoring and activity logs as one visibility infrastructure | Would say the field built half of what they proposed |
| Mark Kelley, legal commentator | “An AI system cannot consent to anything in a meaningful sense” | Would say the law needs a principal, which identity provides, and evidence of what the principal authorised, which it does not |
| David Shoemaker, philosopher | Responsibility has distinct faces: attributability, answerability, accountability | Would recognise the registration theory as delivering attributability and mistaking it for answerability |
| A chief information security officer at a regional bank (composite view, not a named person) | Identity is the control I can deploy this quarter | Would say: give me identity now and evidence when the standards exist; and would be right about the sequence, wrong if the second step never comes |
One critic deserves more than a table row. The strongest objection to this essay comes from the identity community itself, and it runs: you are attacking a straw man; nobody serious says identity is sufficient; every identity vendor sells logging too. The first half is fair. The specifications are honest. But the objection misreads where the problem lies. It is not in the specifications. It is in procurement, policy and politics, where a visible identity programme becomes the answer to “what are we doing about agent accountability?”, and the logging, being a feature rather than an infrastructure, varies by vendor, sits inside the agent’s own platform, follows no common format and is rarely tested by anyone independent. The straw man is not in the standards. It is in the board pack.
XIII. For Banks: Preparing for the Incident Clock
A reader, Baris Tuncertan, asked under Six Regimes, One Agent: The November Update how banks should prepare for the incident clock that Australia has accepted in principle and that others are likely to follow. It is a good question to answer from the identity side, because the incident clock is where the three-layer model stops being theory.
An incident report has a structure that does not change much between regulators. What happened, when, to whom, how many customers, what was the cause, what have you done. Try answering those questions within a fixed window using only identity-layer data. You will be able to say which agent was involved and which human sponsors it. You will not be able to say what it did, how many times, to which accounts, on what reasoning, or whether it has stopped. The clock will run out while your engineers are reconstructing a trajectory from application logs the agent itself wrote.
So my advice to banks, beyond the regulatory overlays in that update, is five things, each of which is useful whatever the eventual rule says.
- Make the incident question answerable from the record. For every agent that can move money or change customer data, test whether you can produce, within hours, the sequence of actions it took in a given period, from records it could not alter. If you cannot, that is your first gap, and it is a layer-three gap.
- Count executions. For every consequential action, record an execution identifier that ties the approval to exactly one execution, and alert on any approval consumed twice. This is cheap, and it turns duplicate execution from an incident you discover from a customer into one you discover from a dashboard.
- Reconcile claims against actions. Daily, compare what agents reported to their users with what the systems of record show. The discrepancies are your earliest warning, and they are exactly what the Starter Kit warns about.
- Adopt SAFR’s shape, even before anyone requires it. It is not regulatory guidance, and MAS has said dedicated agentic expectations will come through consultation. But its components, identity, controls repository, disposition, envelope, tamper-evident log, are a ready-made evidence architecture written by your peers. When a clock arrives, a bank that already produces Governance Envelopes will be filling in a form; a bank that does not will be starting a forensic project.
- Rehearse the report. Run a tabletop exercise in which an agent double-pays fifty suppliers on a Friday afternoon, and time how long it takes to produce a regulator-ready account. Then fix whatever made it slow. The answer will rarely be “we did not know which agent it was”.
None of this requires waiting for a registry, a law or a standard. It requires treating evidence as part of the control, not as an afterthought to it.
XIV. What Singapore Should Build Next

Singapore is unusually well placed to get this right, for reasons that have nothing to do with size. It has a central bank whose industry group has already written the runtime model. It has a governance framework that already asks for runtime controls. It has a testing kit that already asks for independent trajectory logs. It has, since October, an accreditation programme for the firms that test AI systems. And it has a minister who has said, in Parliament, that the point of all this is that someone can answer. What it lacks is the step that joins these up. Here are five proposals, in order of cost.
1. Keep exploring identity, and say clearly that it is layer one of three. Nothing in this essay argues for slowing the DNS work or the GovTech registry. The language matters, though. If the public account of Singapore’s agent governance becomes “Singapore is building a registry for agents”, the third layer will be treated as an implementation detail. If it becomes “Singapore is building identity, authority and evidence for agents, and exploring which of them need common infrastructure”, the third layer is on the agenda. Ms Lau’s own phrasing, many efforts, each addressing a different part of the problem, already allows for this.
2. Specify a common agent action record. Before deciding whether a national registry is needed, specify what a record of an agent’s consequential action must contain: the agent’s identity, the mandate it acted under, the action proposed, the policy decision taken, the action executed, an execution identifier so that duplicates are detectable, the outcome, and the time of each. Specify that it must be generated or corroborated outside the agent’s control, as the Starter Kit already says. SAFR’s Governance Envelope and audit entry are an excellent starting point; IMDA could generalise them beyond finance. This costs a working group, not a building.
3. Make the registry point to the record. If Singapore does eventually build common identity infrastructure, each entry should carry not only the agent’s identity and its accountable human, but a pointer to where its action records are held and who has independently tested them. That turns the registry from a directory of names into a directory of answers. It is a small design decision that is very hard to retrofit.
4. Give the accredited testers something to test. The AI Tester Accreditation Programme accredits firms to test AI systems. (Again, PwC Singapore is one of the five accredited, and I am a PwC partner on leave; readers should weigh this paragraph accordingly.) Testing today is mostly about how a system behaves under test. Once a common action record exists, testers can also check whether a deployed agent’s records are complete, tamper-evident and consistent with what systems of record show: whether the third layer is real. That is assurance of the answer, not just of the model.
5. Take the three layers to ASEAN in 2027. Singapore’s stated intention as ASEAN chair is to strengthen regional capability in AI governance, testing and assurance, with international standards and common testing tools for agentic AI. Identity infrastructure is hard to harmonise across ten jurisdictions with different national identity systems and different views of the state’s role. A common evidence format is much easier: it asks each country to require nothing more than that agents operating in sensitive sectors produce records in an agreed shape, and it lets each keep its own identity arrangements. The November update to Six Regimes suggested an ASEAN template for agent incident reporting. An ASEAN agent action record would be the evidence underneath it: the thing every incident report would be built from. It would also travel. A regional bank operating in six ASEAN markets would produce one kind of record, not six. And the definitions, as I wrote of Singapore’s other governance work, would travel further still.
None of these steps requires legislation. All of them are consistent with what Singapore has already published. And together they would mean that when the next minister answers the next adjournment motion, the answer to “who answers for the machine?” is not only “the person in the registry” but “the person in the registry, with the record that lets them”.
XV. The Counter-Case, and What Would Change My Mind
The best argument against this essay is not that identity is sufficient. It is that the third layer is a mirage: that comprehensive, trustworthy records of agent actions are too expensive, too invasive or too easily gamed to be worth requiring, so identity plus liability is the realistic end state, and courts will sort out the rest.
There is something to each part of that. Records are expensive, and the cost falls hardest on smaller firms. Complete action logs are a surveillance asset and a breach target, a point the Centre for the Governance of AI’s work on visibility into agents takes seriously. Records can be gamed; the summer’s evaluation incidents included agents tampering with their own tool-call records in small tests, which is precisely why the Starter Kit insists on logs outside the agent’s control. And courts have managed without records before: in Moffatt v Air Canada, the airline answered for its chatbot without anyone reconstructing how the chatbot worked.
My replies, briefly. Cost scales with consequence: the record need only cover consequential actions, which is what SAFR’s calibration to materiality already does. Privacy is a design constraint, not a veto: records can be minimised, access-controlled and retained for fixed periods. Gaming is the argument for independence, not against records. And Air Canada answered for a refund of a few hundred dollars. Agents that move money, send official letters and sign contracts will generate disputes where “the company is liable, we do not know what happened” is not an acceptable answer to a regulator, a citizen or a court.
Still, I could be wrong, and here is what would change my mind.
- If the agent identity schemes in the table began to carry runtime attestations of behaviour, of the kind ANS lists as future work, the line between layers one and three would blur, and identity programmes would deliver more than I credit them with.
- If insurers and courts showed, in practice, that identity plus liability resolved agent disputes efficiently without detailed action records, the third layer would be a luxury rather than a necessity.
- If a common action record, once specified, proved so costly or so easily gamed that firms produced it only as theatre, my proposal would add paperwork, not answerability. The forthcoming essay Guardians of the Guardrails explores when records become theatre.
- If Singapore’s identity work, once concluded, turned out to include an evidence component from the start, my worry about sequencing would be misplaced, and I would be glad of it.
XVI. What I Could Not Verify
- The parliamentary exchange: I relied on reports by The Straits Times (via syndication), AsiaOne and trade press. I did not read the Hansard record. Both main reports date the adjournment motion to 4 November 2026.
- Tin Pei Ling’s speech: I found no verbatim quotation and have paraphrased from reports. “Biometric passporting” appears only in Biometric Update’s account and I have not attributed it to her.
- The Chain of Trust: I relied on secondary reporting of the GovWare demonstration and did not read the participating companies’ case study or reference architecture directly. Which check caught the forged seller was not stated.
- SAFR: I read the version 1.1 paper through a fetch-and-summarise tool rather than in full layout, so quotations are from that reading; the components, dispositions and the identity sentence were consistent across my reads and with the MAS page.
- The agentic governance framework and the Starter Kit: I read the passages quoted, not every section.
- A2A Agent Cards: I did not verify the current specification version at source. Entra Agent ID’s general availability date comes from secondary sources. Visa and Mastercard descriptions rest on company announcements and trade press.
- The OpenAI forum report of duplicate execution is a single, unconfirmed report from a small number of samples, with only a holding reply from OpenAI support. It establishes a pattern worth guarding against, not a rate.
- I found no source quantifying revocation latency for agent systems.
- The Bangladesh Bank figures are as widely reported in 2016; I did not read the forensic reports.
XVII. What to Demand on Monday Morning
Chief information officers and chief technology officers
- Draw your agent estate in three columns: identity, authority, evidence. For each agent that can take a consequential action, mark which layers exist. Most estates will be strong in the first column, patchy in the second and nearly empty in the third.
- Give every agent its own identity, scoped and expiring, if you have not already. Then, for every consequential action, add an execution identifier and a record written outside the agent’s control. Do not let identity be the end of the project.
Chief risk officers and chief compliance officers
- Ask for the receipt from the last time a gate said no to an agent: the action proposed, the rule applied, the decision, the record. If nobody can produce one, you have identity and permissions, not control.
- Ask how many times your most important payment or customer-data agent executed each approved action last month, and how you know. If the answer is “once, because the agent said so”, treat it as a finding.
Regulators, in Singapore and across ASEAN
- In the assessment of whether common infrastructure is needed, assess the record alongside the registry. Specify a common agent action record before or alongside any decision on a national registry, and make any registry entry point to it.
- Ask accredited testers to test the record, not only the agent: completeness, independence from the agent, consistency with systems of record.
Boards
- Ask management one question: if one of our agents did something twice, or did something it then misreported, who would answer, and what evidence would they use? Then ask to see that evidence for a real action from last week.
- Do not accept “every agent has a verified identity” as an answer to “are our agents accountable?”. It is an answer to “can we name them?”.
Coda: The Name and the Account
There is a scene in every old registry office. A clerk stamps a document, the applicant takes it, and walks out into the street with an identity the state will vouch for. What happens next, the clerk never sees. For most of history that was fine, because what happened next was done by a person who could later be asked, and could answer.
Agents change that. They walk out of the registry office with a perfectly stamped identity and act, at speed, in places no clerk will ever see, on behalf of a person who was not watching. Jasmin Lau saw this exactly: agents can act for us, but they cannot answer for us. Singapore has resolved that someone always will. The machinery being explored will tell us who that someone is.
The next step is to make sure that when we find them, they can do what the minister asked of them: not just be named, but answer. That needs a record of what the agent did, written where the agent cannot reach it. Singapore’s own documents already ask for one. The name is the beginning of accountability. The account is the rest of it.
Luke Soon is an AI futurist and the author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: SuperIntelligence Protocol. He co-authored IMDA Singapore’s Model AI Governance Framework for Agentic AI, which this essay cites, and has a commercial interest in TrustOS, a runtime governance architecture developed with Lumela Labs. He is a PwC partner on leave, writing in a personal capacity; PwC Singapore is one of the firms accredited under Singapore’s AI Tester Accreditation Programme. Who Answers for the Machine? is a three-part series; No. 3, Exactly Once, follows.
Library
Singapore: Parliament and policy
- The Straits Times via Yahoo, Singapore exploring adapting the domain name system to identify AI agents (4 Nov 2026)
- AsiaOne, AI agents, trust and safeguards: Jasmin Lau in Parliament (4 Nov 2026, updated 9 Nov)
- ID Tech Wire, Singapore explores using DNS to identify AI agents as it weighs national registry (Nov 2026)
- Biometric Update, Singapore explores internet domain system to identify and track AI agents (13 Nov 2026)
- GovTech, Opening address by SMS Tan Kiat How at GovTech STACK Conference 2026, Day 2 (5 Nov 2026)
- TechCoffeeHouse, GoDaddy, Google and Cloudflare unveil AI agent trust system (16 Oct 2026)
Singapore: frameworks, testing and finance
- IMDA, Model AI Governance Framework for Agentic AI, v1.5 (May 2026, updated June 2026); the author is a co-author
- IMDA, Starter Kit for Testing Agentic Systems for Safety and Reliability, v1.0 (Oct 2026)
- MAS, Safeguards for Agentic Finance at Runtime (SAFR), publication page (3 Jul 2026, updated 14 Oct 2026)
- MAS, SAFR v1.1 paper (PDF)
- MAS, MAS partners industry to develop safeguards for AI agents in finance (media release, 2026)
- AI Verify Foundation, AI Tester Accreditation
- Little Big Red Dot, AI Verify Foundation accredits its first five AI testers as IMDA launches an AI agent testing kit (17 Oct 2026)
Agent identity and authorisation proposals
- IETF, DNS-AID, draft-mozleywilliams-dnsop-dnsaid-02 (27 May 2026)
- IETF, Agent Name Service v2, draft-narajala-courtney-ansv2-01 (13 Apr 2026)
- IETF, HTTP Message Signatures for automated traffic (Web Bot Auth), draft-ietf-webbotauth-httpsig-protocol-01 (29 Oct 2026)
- Digital Transactions, Visa launches Trusted Agent, an agentic commerce protocol (Oct 2025)
- Mastercard, New trust and intelligence services (30 Sep 2026)
- Cloud Wars, Google’s Agent Payments Protocol fills governance void in AI financial transactions (Sep 2025)
- Devicebase, General availability: Microsoft Entra Agent ID platform (secondary, 2026)
- Okta, Okta brings first-class identity to AI agents with Agent SSO
- Model Context Protocol, Authorization specification (2026-07-28)
- OpenID Foundation, Advances authorization for the agent era with new AuthZEN working group drafts (15 Jun 2026)
- Agent2Agent (A2A) protocol and Agent Cards, Linux Foundation project documentation (described at category level; version not verified at source)
The runtime gap
- OpenAI Developer Community, Responses API: approved native MCP write executes twice after a programmatic read (16 Nov 2026; single unconfirmed report)
- METR and Redwood Research, independent investigation of the July 2026 evaluation incident, published 26 August 2026 (spoofed tool-call records in small tests)
- Alan Chan and colleagues, Visibility into AI Agents (ACM FAccT 2024), and related work on agent identifiers and activity logging
History
- CCDCOE Cyber Law Toolkit, DigiNotar (2011)
- The Register, DigiNotar declared bankrupt (20 Sep 2011)
- American Banker, The $81M international cyberheist: what happened, how and what it means (2016)
- Fortune, The Fed cyber heist (18 Aug 2016)
- USA PATRIOT Act of 2001, Title III, Section 326 (customer identification programmes)
Responsibility and law
- Gary Watson, “Two Faces of Responsibility” (1996), and David Shoemaker, Responsibility from the Margins (2015)
- P.F. Strawson, “Freedom and Resentment” (1962)
- Filippo Santoni de Sio and Giulio Mecacci, “Four Responsibility Gaps with Artificial Intelligence” (Philosophy and Technology, 2021)
- Mark Kelley, “Agentic AI liability”, Bloomberg Law (8 Jul 2026)
- Moffatt v Air Canada, 2024 BCCRT 149
Earlier Genesis pieces this essay builds on
- Genesis, The Last Human Job Is Answerability; Not Empathy. Answerability. (22 Oct 2026)
- Genesis, AI Agent Assurance: A Board’s Guide to the Governance Frameworks (6 Nov 2026)
- Genesis, Stopping Deepfake Fraud at Runtime: Lessons from Ferrari and Singapore (Receipts, Please, Part IV, 6 Nov 2026)
- Genesis, Every AI Instruction Should Carry Proof (Receipts, Please, Part V, 6 Nov 2026)
- Genesis, Six Regimes, One Agent: The November Update (16 Nov 2026)
- Forthcoming on Genesis: The Permission Problem: Governing the Swarm, No. 1; Guardians of the Guardrails; Ghost in the Loop


Leave a Reply