ReAct

ReAct (Reasoning + Acting) is the foundational agent loop introduced in 2022. At each step, the agent reasons out loud about what to do next, takes an action (calls a tool), observes the result, then reasons again. Nearly every modern agent uses some variant.

The loop

Thought: I need to find X. I'll search for it.
Action: web_search(query="X")
Observation: [search results]
Thought: Based on the results, I need to...
Action: [next tool call]
Observation: [result]
...
Thought: I have enough. Final answer: ...

Why reasoning out loud helps

Having the LLM write its reasoning explicitly before each action improves quality for complex tasks. The model commits to a rationale, which makes its actions more coherent. And the reasoning is in the context for subsequent steps.

Modern tool-use APIs (Claude, GPT-4, Gemini) have native support for ReAct-like reasoning via structured output, where reasoning and tool calls are emitted in a structured format the orchestrator parses.

Stop conditions

ReAct loops need clear stop conditions:

• Agent emits a final answer (not a tool call)
• Max step count reached (typically 10-30)
• Max cost budget reached
• Timeout reached
• Infinite-loop detection (same tool, same args, N times)

Common failure modes

• Tool hallucination, LLM calls a tool that doesn't exist or with wrong args
• Premature answer, agent stops exploring when it should keep going
• Infinite loop, agent repeats the same reasoning + action forever
• Lost-in-context, after many steps, the context is long and the model starts forgetting earlier reasoning

Variants

• Plan-then-execute, agent plans full sequence first, then executes. Works for tasks with known structure.
• Reflection, agent reviews its own trace and critiques itself before continuing. Higher quality, higher cost.
• Multi-agent ReAct, agents call other agents as tools.

When ReAct is the wrong shape

For tasks with a known execution plan, skip ReAct and use a deterministic workflow. ReAct's value is in exploration; if the path is known, the reasoning overhead is waste.