ReAct (Reason → Act → Observe Loop)

What Problem It Solves

If the next step depends on observations (tool outputs), you need a control loop:

• decide what to do next
• call tools
• update context
• repeat until done

When to Use

• Unknown number of tool calls.
• The environment is interactive (search, APIs, file ops).
• You need a “stop condition” (final answer).

Core Flow (Action Schema)

flowchart TD
  S["State (messages)"] --> D["Decide action (JSON)"]
  D -->|tool| T["Call tool"]
  T --> O["Observation"]
  O --> S
  D -->|final| F["Final answer"]

Evolution Path

• Built on: Tool calling + Structured output + Loop controller
• Specializations:
• Agentic RAG = ReAct + retrieval tool + evidence ledger
• Governance hooks = policy/guardrails/HITL around tool calls

Repo Reference

• Code: src/agent_patterns_lab/patterns/react.py
• Example: examples/21_react_loop.py
• Tests: tests/test_react.py