ReAct Pattern

ReAct (Reason + Act) is an agent pattern that interleaves intermediate reasoning with tool use. The model writes a thought, chooses an action, observes the result, and repeats.

When to use

• Tasks needing adaptive tool use
• Debugging-friendly traces (thoughts + actions)
• Single-agent problems without heavy multi-actor coordination

Structure

Thought: ...
Action: tool_name(args)
Observation: ...
Thought: ...
Action: finish(answer)

In modern stacks, “Thought” may be internal, and actions are structured function calls—but the loop is the same.

Implementation notes

• Cap iterations
• Feed tool errors back as observations
• Prefer structured actions over free-text tool DSL when possible
• Don’t require verbose thoughts in production if latency/cost dominate—keep the loop

Failure modes

• Verbose reasoning without action
• Ignoring observations
• Looping on the same failed action

Evaluation tips

Grade both final answers and whether observations influenced subsequent actions (no “ignore the tool result” behavior).