The ReAct Loop

Intuition

ReAct (Yao et al., 2022) lets a language model alternate between thinking, acting (calling a tool) and observing the result, so it can look things up instead of guessing. The agent code is just a loop that shuttles messages between the model and the tools.

Precise explanation

Transcript as a list of messages {"role": ..., "content": ...}:

Start with {"role": "user", "content": question}.
Repeat up to max_steps:
- text = model(transcript) (our scripted model is a Python function); append {"role": "assistant", "content": text}.
- Parse it (AG-02). Final → return the answer. Tool call → dispatch (AG-01) and append {"role": "tool", "content": json.dumps(result)}. Invalid → append {"role": "tool", "content": json.dumps({"ok": False, "error": ...})} so the model can fix itself.
Out of steps → status "step_limit".

sequenceDiagram participant U as user participant M as model participant T as tools U->>M: question M->>T: tool_call(search, …) T->>M: observation M->>U: FINAL: answer

Worked example

Scripted model: first reply calls add(2, 3); after seeing {"ok": true, "result": 5} it replies FINAL: 5. The loop returns answer "5" after 2 model calls.

Common mistakes

Not appending observations (the model never sees results).
Crashing on invalid output instead of reporting the error.
No step limit.

Why this matters

This loop — model, tools, observations — is the core of tool-using LLM agents.

References

ReAct: Synergizing Reasoning and Acting in Language Models — Shunyu Yao et al. (S036)
Building effective AI agents — Anthropic (S037)