Recipe
ReAct Agent Pattern
Reason + Act — the foundational loop for tool-calling agents that think before they execute.
ReAct interleaves reasoning traces with action steps. The model generates a thought, decides which tool to invoke, observes the result, and repeats until it can produce a final answer. This grounds every action in explicit chain-of-thought.
The Loop
- Thought — analyze the query and plan the next step
- Action — select and invoke a tool with structured input
- Observation — ingest the tool output into context
- Repeat — loop until a final answer is ready
Why it matters
Without reasoning traces, tool-calling agents become brittle — they invoke tools blindly and struggle to recover from errors. ReAct makes the agent's decision process transparent, debuggable, and self-correcting. Every observation feeds back into the next thought, creating a tight feedback loop.
When to use it
- Multi-step research tasks that require web search, code execution, or database queries
- Agents that must verify their own outputs before responding
- Any workflow where tool selection depends on intermediate results
Meridian tip: Combine ReAct with structured output schemas to enforce valid tool calls on every action step.