What is it?
Chain-of-thought (and extended / interleaved thinking)
Chain-of-thought means prompting a model to write out its reasoning step by step before giving a final answer, which tends to improve accuracy on hard problems. "Extended thinking" is a controllable version where the model writes out a plan before acting. "Interleaved thinking" is when the model thinks again after receiving a tool result, to evaluate what it got and decide the next step. Think of it as the agent's scratchpad.
From the AI agent glossary
As the Planning and Reasoning glossary puts it
Chain of thought is the technique of having a model write out its intermediate reasoning steps before giving a final answer, instead of jumping straight to the result (Chapter 1). The written steps act as scratch paper: each one becomes part of what the model reads next, giving it room to work through a problem it could not solve in a single leap.
Chain of thought is the seed from which most of modern reasoning grew. The longer, self-checking reasoning of advanced models like DeepSeek-R1 (Chapter 4) is essentially a chain of thought that the model learned to produce and extend on its own.
Where Chain-of-thought is explained
- Chain of Thought, Teaching Models to Think Out LoudPlanning and Reasoning
- Start Here: Planning and Reasoning, Explained SimplyPlanning and Reasoning · How to use these explanations
- Code Review and Security AgentsBuilding Coding Agents and Harnesses · Cursor's security agents and the 15-line prompt
- ReAct, Reasoning and Acting TogetherPlanning and Reasoning · Two failure modes ReAct fixes