🦸🏻#12: How Do Agents Learn from Their Own Mistakes? The Role of Reflection in AI

Last week, we explored how reasoning and planning work together to make AI more effective – allowing models not just to think, but to structure their thoughts into goal-driven actions. Yet, even with strong reasoning and planning, AI still lacks something essential: the ability to learn from its own decisions.

That’s where Reflection comes in. More than just thinking ahead, agentic AI needs to analyze past actions, recognize mistakes, and refine its strategies – just like humans do (sometimes). Without reflection, AI risks repeating errors instead of improving. Andrew Ng sees Reflection as a key design pattern for agentic AI, enabling models to critique and refine their own outputs for better decision-making.

Today, we dive into Reflection as a core building block of agentic AI, exploring how frameworks like Reflexion and ReAct (and others) enable self-assessment and iterative learning. As AI moves toward greater autonomy, self-reflection is emerging as a critical capability – closing the loop between action and continuous learning.

What’s in today’s episode?

• Philosophical Roots

• Understanding Reflection in AI

• Implementation Details: Reflexion and ReAct Frameworks

  • Reflexion: Learning via Self-Feedback

  • ReAct: Interleaving Reasoning and Acting

  • Impact on Decision-Making

• Other Reflection-Based Approaches and Innovations

• Future Directions and Emerging Trends

• Concluding Thoughts

• Resources that were used to write this article (we put all the links in that section)

_{We apologize for the anthropomorphizing terms scattered throughout this article – let’s agree they are all in ““.}

Philosophical Roots

The ability to reflect – to analyze one’s own thoughts and actions – has long been recognized as fundamental to intelligence. Socrates championed the practice of questioning one’s own beliefs, arguing that only through introspection can we separate sound reasoning from flawed assumptions. Ancient Eastern philosophy echoed this idea, with Confucius placing reflection above both imitation and experience as the noblest path to wisdom. Throughout history, reflection has been seen as the mechanism that sharpens judgment, refines decision-making, and fuels personal and intellectual growth.

From the Stoics, who practiced nightly self-examination, to Descartes, who probed the very nature of thought itself, reflection has been a core theme in philosophy. Thinkers like Aristotle and Kant drew distinctions between contemplation and action, recognizing that purposeful deliberation is key to meaningful decision-making. In more recent times, John Dewey described reflective thought as the careful and persistent evaluation of beliefs in light of evidence, enabling individuals to act with foresight rather than impulse.

Donald Schön later expanded on this by distinguishing between reflection-in-action – adjusting and adapting in real time – and reflection-on-action, where individuals analyze past decisions to improve future ones. This idea reinforced the notion that expert decision-making isn’t just about planning but also about dynamically evaluating and refining actions as they unfold. These concepts have profoundly influenced fields from cognitive science to education, shaping our understanding of how learning, reasoning, and action must work together to drive true intelligence.

Today, these ideas are finding their way into AI.

Understanding Reflection in AI

In the context of AI agents, reflection refers to an agent’s ability to think about its own actions and results in order to self-correct and improve. It is essentially the AI analog of human introspection or “System 2” deliberative thinking​. Instead of merely reacting instinctively (like a fast, heuristic System 1 response), a reflective AI will pause to analyze what it has done, identify errors or suboptimal steps, and adjust its strategy. (the concepts of System 1 and System 2 thinking were proposed by Daniel Kahneman in his book “Thinking, Fast and Slow”). This mechanism allows an AI agent to learn from its experiences without new external training data, by internally evaluating feedback. Through self-reflection, an agent can correct previous mistakes and generate improved solutions on the next attempt, embodying a form of self-improvement loop.

AI pioneer Andrew Ng sees Reflection as a core component of agentic AI, alongside Planning, Tool Use, and Multi-agent Collaboration – the building blocks that could define the next wave of AI progress. Rather than just generating answers, reflective AI models critique and refine their own outputs, identifying flaws, making improvements, and iterating until they reach a stronger result.

This self-review loop – generate → critique → improve – has already shown major performance gains across tasks like coding, writing, and question-answering. Ng emphasizes that Reflection enables AI to self-correct without always needing external feedback, making systems more reliable and autonomous. 

Reflection in AI does not stand alone – it interplays with other core components of an agentic system. As we wrote before, an autonomous agent typically has several building blocks working in concert​. First is Profiling, where the agent is given a role or objective that frames its behavior (defining its identity, goals, and constraints)​. Next comes Knowledge, which provides the agent’s base information or access to facts (such as a knowledge base or pretrained model of the world). The agent also maintains Memory – storing context from past interactions or steps (both short-term, like the current conversation or trajectory, and long-term, like accumulated experience)​. With knowledge and memory, the agent engages in Reasoning and Planning, breaking down tasks, making inferences, and formulating an action plan​. It then takes Actions to execute the plan (e.g. calling tools, producing outputs, or moving in an environment)​. Finally comes Reflection: the agent evaluates the outcomes of those actions against the goals, learning from any mistakes or unexpected results​. This reflective step “evaluates outcomes to influence future reasoning and behavior,” closing the feedback loop in the agent’s workflow​.

Crucially, reflection allows the agent to adjust itself dynamically. In a well-designed agentic system, this means the operational flow is cyclic: profiling defines the role, knowledge provides context, memory and reasoning guide an action, the agent acts, and then reflection assesses how that action went​. Any lessons learned (e.g. an action failed or a reasoning path led nowhere) are fed back into the agent’s memory or planning module to inform the next cycle. Over time, this leads to better performance as the agent accumulates reflective insights. This idea of continuous self-improvement via reflection has strong theoretical appeal – it’s essentially a form of on-the-fly adaptation. Notably, it does not require retraining the model’s weights each time; instead, the learning happens at the knowledge and planning level through natural language or symbolic feedback. Researchers have even likened this to a kind of “verbal reinforcement learning” for language agents, where the model reinforces good behaviors through linguistic feedback rather than gradient updates​.

In summary, reflection in AI provides a mechanism for an agent to meta-reason about its own reasoning, enabling a higher level of autonomy.

Implementation Details: Reflexion and ReAct Frameworks