Demystifying the Top 20 Agentic AI Concepts:

By Dr Luke Soon

October 21, 2025

As a researcher and consultant specialising in artificial intelligence applications within the financial sector, I’ve long been fascinated by the evolution of AI from passive tools to proactive entities. Agentic AI—systems that exhibit autonomy, reasoning, and adaptability—represents a paradigm shift, particularly in an industry as regulated and data-intensive as finance. Drawing from the insightful infographic circulating in AI communities, which distils these concepts into accessible visuals, this post delves into the top 20 agentic AI concepts. I’ll explain each technically, grounded in their foundational principles, and illustrate their transformative potential in financial services with real-world examples and citations. My aim? To equip finance professionals with a rigorous yet practical understanding, enabling them to harness these technologies for enhanced efficiency, risk management, and innovation.

In finance, where decisions must balance speed, accuracy, and compliance, agentic AI promises to automate complex workflows while augmenting human expertise. From fraud detection to portfolio optimisation, these concepts are not mere abstractions; they are deployable architectures. Let’s explore them systematically.

Core Foundations: Agents, Brains, and Memory

1. Agent

At its essence, an agent is an autonomous software entity capable of perceiving its environment, reasoning about goals, and executing actions to achieve them. Technically, agents operate within a feedback loop: observation → deliberation → action → evaluation. In reinforcement learning terms, this mirrors a Markov Decision Process (MDP), where states represent market conditions, actions denote trades, and rewards quantify profitability.

In financial services, agents streamline customer onboarding by autonomously verifying identities and assessing creditworthiness. For instance, JPMorgan Chase has deployed AI agents to reduce onboarding time by 40%, integrating KYC (Know Your Customer) checks with real-time data pulls. Similarly, HSBC’s agentic systems handle trade settlements, minimising errors in high-volume transactions.

2. LLM (Large Language Model)

The LLM serves as the cognitive core, a transformer-based neural network trained on vast corpora to generate human-like text. It processes inputs via attention mechanisms, encoding semantic relationships to infer intent or generate plans. Fine-tuning with domain-specific data, such as SEC filings, enhances precision.

Financial applications abound: Goldman Sachs leverages LLMs for sentiment analysis on earnings calls, predicting stock movements with 85% accuracy. Barclays uses them in chatbots for personalised advice, reducing query resolution time by 60%.

3. Planning

Planning involves hierarchical task structuring, often using algorithms like Monte Carlo Tree Search (MCTS) or classical planners (e.g., STRIPS) to decompose objectives into subgoals. Agents forecast outcomes via simulation, optimising for constraints like regulatory limits.

In trading, planning agents at Citadel simulate portfolio rebalances, anticipating volatility spikes. Deloitte highlights how banks use planning for loan origination, sequencing credit checks and risk assessments to cut processing delays.

4. Memory

Memory in agents comprises short-term (contextual buffers) and long-term (vector databases like FAISS) stores, enabling episodic recall. Techniques like experience replay from deep learning allow agents to learn from historical interactions without catastrophic forgetting.

For compliance, memory-equipped agents at Standard Chartered track transaction histories, flagging anomalies against AML (Anti-Money Laundering) patterns. In wealth management, Fidelity’s systems recall client preferences for tailored recommendations.

Execution and Adaptation: Tools, Decomposition, and Reflection

5. Tool Use

Tool use extends agent capabilities via APIs or external modules, invoked through function calling in LLMs. This hybrid architecture—e.g., ReAct (Reason + Act)—alternates reasoning with tool queries, ensuring grounded outputs.

In fraud detection, agents at PayPal integrate with blockchain APIs to verify cross-border payments in real-time. Moody’s agents use market data tools for dynamic credit scoring.

6. Task Decomposition

This breaks monolithic tasks into atomic subtasks using recursive prompting or graph-based planning (e.g., AND-OR trees). It mitigates LLM hallucination by localising complexity.

Commercial banks apply decomposition for KYC workflows: subtasks include document extraction, entity resolution, and sanction screening, as seen in BBVA’s automation suite.

7. Self-Reflection

Self-reflection employs meta-cognitive loops, where agents critique outputs via another LLM instance, scoring for coherence or accuracy. This draws from reflective equilibrium in decision theory.

In auditing, KPMG’s agents self-review tax filings, reducing errors by 30% through iterative refinement. SymphonyAI notes its use in financial crime prevention for ongoing model validation.

8. Autonomous Execution

Full autonomy minimises human intervention via closed-loop control, with safeguards like human-in-the-loop (HITL) for high-stakes actions. Technically, it’s event-driven, using pub-sub patterns for reactivity.

NVIDIA-powered agents at Deutsche Bank execute trades autonomously during off-hours, enhancing liquidity management. Workday’s compliance agents monitor regulations 24/7.

Knowledge and Collaboration: Context, Retrieval, and Teams

9. Context Window

The context window is the token limit (e.g., 128k in GPT-4o) for input processing, managed via sliding windows or compression techniques like summarisation to avoid truncation.

In financial reporting, agents at Ernst & Young use expanded windows to analyse full annual reports without losing fiscal nuances.

10. Retrieval-Augmented Generation (RAG)

RAG fuses dense retrieval (e.g., DPR) with generation, querying vector stores for relevant documents before synthesis. It combats hallucinations by grounding responses in external knowledge.

Hatchworks details RAG’s role in personalised financial advising at Capital One, pulling from proprietary datasets for accurate projections. RavenPack’s finance-specific RAG analyses unstructured news for trading signals. Lumenova AI applies it to regulatory updates, ensuring compliance queries are fact-based.

11. Multi-Agent Systems

These orchestrate specialised agents via protocols like message passing in MAS (Multi-Agent Systems), enabling emergent intelligence through negotiation or consensus.

SmythOS describes multi-agent setups at BlackRock for portfolio optimisation, where risk and yield agents collaborate. Deloitte’s M&A agents use sequential patterns for due diligence. Akira AI’s compliance systems distribute tasks across monitoring agents.

12. Reinforcement Learning

RL trains agents via policy optimisation (e.g., PPO algorithm), maximising cumulative rewards in stochastic environments like markets.

arXiv papers showcase RL at Jane Street for high-frequency trading, adapting to volatility. Neptune.ai lists applications in robo-advisory at Vanguard, yielding superior ROI. Coursera’s strategies optimise order execution at Citadel.

Reasoning and Guidance: Chains, Thoughts, and Goals

13. Prompt Chaining

Prompt chaining sequences interdependent queries, propagating outputs as inputs, akin to a directed acyclic graph (DAG) for workflow execution.

In financial forecasting, CFI uses chaining for scenario analysis, linking macroeconomic prompts to balance sheet projections.

14. Inner Monologue

This internal verbalisation simulates stream-of-consciousness reasoning, captured via hidden states in LLMs to refine latent representations.

AWS employs it in hypothesis testing for credit risk at Wells Fargo.

15. Action Space

The action space defines discrete or continuous options (e.g., buy/sell/hold in trading), pruned via pruning heuristics to reduce computational load.

In options pricing, agents at Morgan Stanley explore vast spaces for arbitrage opportunities.

16. Goal-Oriented Behaviour

Agents pursue objectives via utility functions, using A* search or gradient-based optimisation to navigate multi-step paths resiliently.

Domo’s guide illustrates goal agents in risk response at Santander, prioritising threats dynamically.

Resilience and Orchestration: Healing, Selection, and Integration

17. Self-Healing Agents

These detect anomalies via sentinel monitors and remediate via rollback or reconfiguration, inspired by chaos engineering.

Trullion’s accounting agents auto-correct ledger discrepancies at PwC. IBM’s fraud agents self-adjust thresholds.

18. Dynamic Tool Selection

Selection uses contextual bandits or meta-learning to choose tools, balancing exploration with exploitation.

In wealth management, xCube LABS agents select analytics tools for client queries.

19. Chain-of-Thought

CoT elicits step-by-step reasoning via prompted intermediates, boosting accuracy on arithmetic or logical tasks by 20-50%.

YouAccel’s framework applies CoT to valuation models at Deloitte, enhancing predictive fidelity. arXiv’s AD-FCoT uses analogies for market forecasting. Warp News demonstrates it for earnings analysis.

20. Orchestration

Orchestration coordinates components via frameworks like LangChain or AutoGen, managing state and fault tolerance in distributed systems.

AWS Marketplace agents orchestrate mortgage refinancing at Quicken Loans. McKinsey’s blueprint integrates it enterprise-wide at BNP Paribas.

Conclusion: Agentic Future in Finance

These 20 concepts form the scaffolding for next-generation financial AI, from solitary agents to collaborative ecosystems. As we’ve seen, their adoption—evident in deployments by firms like JPMorgan, BlackRock, and HSBC—yields tangible gains in autonomy and precision. Yet, challenges remain: ethical governance, bias mitigation, and regulatory alignment. I urge financial leaders to pilot these in sandboxes, starting with low-risk areas like reporting.

For those eager to dive deeper, explore the cited resources or reach out—I’m always keen to discuss implementations. Agentic AI isn’t just transformative; it’s essential for a resilient financial landscape.