AI Agent Architecture — ReAct, Memory, Planning and Multi-Agent Systems · Reasoning — How an LLM Thinks Before Acting
When Reasoning Fails — CoT Pitfalls and Limits of Self-Correction
Reasoning — How an LLM Thinks Before Acting
Introduction
Chain-of-Thought and Tree of Thoughts are not infallible — models make characteristic errors: they fabricate convincing rationale for wrong conclusions, fall victim to composition errors, struggle with negation, and cannot effectively correct their own mistakes without an external signal. This lesson catalogs the most important CoT pitfalls, the phenomenon of "unfaithful reasoning", and the limits of self-correction — so developers know when reasoning is a risky tool.