ReAct

ReAct — short for 'Reason + Act' — is a prompting pattern, described in published research, where instead of asking a model to just pick an action, you ask it to write out a short thought about what it's trying to do and why, then take an action, then read the observation that comes back, then write another thought, and so on. The explicit 'Thought' step is the distinguishing feature: it gives the model a place to reason about whether its last action worked, whether the plan needs to change, and what to try next, in text the agent (and a person debugging it) can actually read.

Before ReAct, one common approach was to have a model reason through a problem purely in text (chain-of-thought) without taking any real actions, or to have it take actions without showing its reasoning at all. ReAct combined both: reasoning traces interleaved with real tool calls and their real results, so the model's plan can be grounded in what actually happened rather than what it assumed would happen.

Because the intermediate 'Thought' text is visible, ReAct-style traces are also useful for debugging: when an agent goes down the wrong path, the thought steps usually show where its reasoning diverged from what actually happened, which is harder to diagnose from action-only logs.

How it works

The loop shown in the diagram — Thought, Action, Observation — repeats: the model writes a thought about what to do next, takes the corresponding action (typically a tool call), reads the observation that comes back, and writes the next thought based on it. This is a specific way of structuring the 'plan' step inside the broader agent-loop — many implementations of the agent loop use something ReAct-like for that step, though it's one technique among several, not the only way to do it.

Example

Thought: I need to find out why the build is failing before I can fix it.
Action: read_file("ci_log.txt")
Observation: Error: dependency 'foo' version mismatch.
Thought: The log points to a dependency version issue, so I should check the lockfile.
Action: read_file("package-lock.json")
Observation: foo is pinned to an older version than the code expects.
Thought: I should update the pin and rerun the tests.
Action: edit_file("package.json", ...)

How it differs

Common misconceptions

FAQ