No Code, Real Power: Exploring Agentic AI Platforms.

Right then, let’s have a proper look at this agentic workflow business, shall we? It’s a fascinating area, and these sources give us a good grounding.

Unpacking Agentic Workflows: A Step-by-Step Affair

At its core, an agentic workflow is about defining what you actually do when you do what you do. Professor John von Neumann, bless him, showed us the way back in 1948 with his flowcharts, which laid out step-by-step procedures – originally for algorithms, but the principle applies more broadly. Think of it as a recipe. This structured approach isn’t just for computers; it can happen on a social level too.

You see these flowcharts everywhere in the business world. Take a mining company for example, where all sorts of processes can be meticulously mapped out. Even the famously socially awkward Sheldon Cooper from that television programme devised a friendship algorithm – a step-by-step procedure for making chums.

The real power comes when we apply this to AI agents. The idea is to teach these agents by converting standard operating procedures into this workflow logic. One example showed a student taking a company’s donation policy (usually a confusing chunk of plain English) and turning it into a clear decision chart. This allows an AI agent to decide whether a donation can be kept or not, based on specific criteria.

Even something as complex as laws can be broken down this way. Laws are essentially a “code of law” – a systematic collection of statutes. Often written in hundreds of pages of impenetrable language, a student demonstrated how a specific aspect, like defining a “dependent” in tax law, can be visualised as a flowchart. This is essentially what tax software already does – following a strict rulebook. Importantly, in such workflows, AI agents shouldn’t have autonomy; they must stick to the defined procedure.

The Autonomous Side of AI Agents

However, one of the most exciting aspects of AI agents is their ability to make autonomous decisions. In contrast to rigidly defined workflows, here the AI itself can decide on a higher-level algorithm or workflow to achieve a task. They have a choice in how to solve problems based on a degree of intelligence.

A prime example is an early “AI agent” that allowed you to take a picture of your fridge and ask what you could cook. This involves several steps: processing the image, recognising ingredients, and then generating a recipe. Interestingly, when different large language models (LLMs) were asked to outline a workflow for this, they came up with different approaches. Claude AI included steps like checking the freshness of ingredients and a feedback loop for the user to refine the recipe. Mistral AI produced a more linear workflow, perhaps less concerned with user preferences. This highlights how different AI agents can interpret the same task in varying ways.

It turns out that many cooking AI agents don’t have an interactive refinement loop due to the complexity of flavour profiles and maintaining recipe coherence. This shows that sometimes, designers deliberately choose to limit autonomous decisions to manage complexity.

Workflows, Algorithms, and the Fabric of Society

Thinking more deeply, these structured workflows are very similar to algorithms. The technical definition of an algorithm is a set of unambiguous, executable steps that reach a termination point. You can almost think of an algorithm as a recipe – like baking a cake step by step.

A workflow is essentially the same thing, but at a higher level of abstraction, often operating on a social level. Learning to play an instrument, following a profession’s routines, even cultural habits are akin to these if-then algorithmic statements. For instance, driving on the left in London versus the right in California, or greetings in different countries. Even laws can be seen as rules for living at a hard-coded level. Sociologists have various terms for these societal workflows, like habitus or structuration. Ultimately, whether we call them workflows, cultural norms, or laws, they are fundamentally similar to algorithms at different levels.

Putting AI Agents to Work: Use Cases Galore

The sources provide numerous use cases for AI agents and agentic workflows:

•Generating Recipes: As discussed, taking a picture of fridge contents and creating a custom recipe.

•Automating Standard Procedures: Converting donation policies into decision-making processes.

•Navigating Complex Regulations: Helping users understand university administrative rules by searching through vast databases.

•Scientific Discovery: Automating the workflow of a scientist to come up with and test hypotheses.

•Data Exploration: LLMs can assist in brainstorming data set generation, summarising information from websites and data sets, identifying missing values, and even writing and running code for data analysis.

•Content Generation with Specific Formats: LLMs can be prompted to create meeting agendas following a template, summarise reports in a specific number of bullet points, or even generate entire web pages with code.

Teams of AI: The Power of Multi-Agent Systems

For tackling really complex problems, we can employ multi-agent systems – teams of AI agents working together. This “divide and conquer” approach allows for specialisation and the balancing of different perspectives, such as creativity and truthfulness, or accuracy and potential bias. The whole becomes more than the sum of its parts, exhibiting emergent properties.

When setting up these multi-agent systems, the network of connections between agents is crucial. A simple, unmanaged network can quickly become a “spaghetti graph”, with connections increasing exponentially with each added agent, leading to complexity and potential failures. The mathematical solution is often a “hub-and-spoke” network, where a central agent orchestrates the others, leading to a linear increase in complexity, much more manageable and efficient. Think of major airport hubs like Houston or LAX – they facilitate efficient redistribution.

Keeping Humans in the Loop

While AI agents can operate autonomously, it’s often beneficial or necessary to have a human in the loop. This provides a check on the AI’s decisions and ensures quality. The paradigm has even evolved to the point where agents can subcontract humans for tasks they can’t handle themselves. This highlights the complementary strengths of humans and AI.

The Art of Talking to AI: Prompting Matters

Effective communication with these AI agents, known as prompting, is paramount. Instead of just throwing a question at an LLM (known as zero-shot prompting), we can significantly improve the output by specifying several things, drawing on the communication framework of a sender, message, and receiver.

Key aspects of prompting include:

•Defining the Role (Sender): Telling the LLM who it is or to act as a certain persona greatly influences the response.

•Specifying the Output Format (Message): Asking for the output in a specific way (e.g., three bullet points, a table, a short essay) gives you what you literally want.

•Identifying the Audience (Receiver): Explaining who the LLM is addressing allows it to tailor its language and complexity. Explaining something to a five-year-old requires a different approach than explaining it to an expert. Importantly, when building AI agents for wider use, considering what the users might do and putting guardrails in place is crucial to prevent unintended or harmful actions.

•Reasoning Process (How to Pack the Message): Techniques like chain-of-thought prompting, where you break down a task into sub-steps, encourage the LLM to reason more effectively, leading to better results. Asking the AI to first analyse the task or even ask clarifying questions can also be beneficial.

•Style and Perspective: Specifying the desired tone (e.g., funny, serious) can further refine the output.

•Iterative Refinement: Going back and forth with the AI, refining the prompt based on the responses, can lead to much better outcomes.

Studies suggest that by specifying these different aspects, you can significantly improve the quality of the output.

Cumulative Benefits and the Future: Web 3.0

It’s important to understand that the benefits of these approaches are cumulative. Even older foundational models can outperform newer ones if they are placed within well-designed agentic workflows. This suggests that context and structure are incredibly valuable.

Looking ahead, some envision a Web 3.0 where we have a vast diversity of interactive AI agents, much like the app stores of today. This future sees a shift from static web pages (Web 1.0) and dynamic social media apps (Web 2.0) to interactive AI agents that can search, share, and generate.

Addressing the Worries

Finally, it’s worth noting that many of the critiques and supposed limitations of AI often stem from fears and a sense of defensiveness as humans adapt to these intelligent systems. Some criticisms are simply incorrect, while others depend on perspective. The key is to approach this with a critical but also calm and pragmatic mindset – “keep calm and carry on,” as they say.

In conclusion, agentic workflows, whether strictly defined or allowing for autonomy, along with the power of multi-agent systems and the crucial role of effective prompting, represent a significant evolution in how we interact with and leverage artificial intelligence. From automating simple tasks to tackling complex scientific challenges, the potential is immense, and understanding these fundamental concepts is key to navigating this rapidly evolving landscape.