ReAct: Synergizing Reasoning and Acting in Language Models
ReadInterleaves reasoning traces with tool actions so a model can plan, act, and revise. Most agent loops are a variation on this pattern.
My notes
Chain-of-thought gives reasoning without grounding; action-only agents act without a plan. ReAct is just "do both, alternately" — and it is still the skeleton of every agent I build.
What I Use From It
- Reasoning traces make agent failures debuggable — you can read why it chose a tool.
- Grounding reduces hallucination versus CoT alone on HotpotQA / FEVER, but it can get stuck repeating the same action. Always cap steps.
- Best results came from combining ReAct with CoT-self-consistency and switching between them — a hint that routing between strategies beats one fixed loop.
In practice: guardrails, step limits and a trace viewer matter more than the prompt wording.