ReAct

What is it?

ReAct alternates a short decision rationale with actions and observations. The model proposes a tool call; the application runs the tool and carries the real result into the next call. A search result imagined in a chat does not count as an observation.

When does it help?

For work where you cannot decide all the steps at the start, and the direction changes based on a result from a document or environment.

Examples

The situations and responses below are fictional teaching examples; they are not results from a model, tool or benchmark that was actually run.

Simple

Situation

You want to find out whether a museum is open on Monday.

Prompt

User task: Is the fictional City Museum open on Monday?
Application: There is only a read-only museum_hours(name) tool. At most 2 calls; if there is no tool, say so and stop.
Model: Write the required action and a one-sentence reason. Do not say open/closed before the application's result arrives.
Representative tool result: {name:"City Museum", monday:"closed", source:"official hours record", updated:"2026-09-01"}
Next model call: Take the task and this real tool result; answer while stating the record's date, and finish if it is sufficient.

Sample output

Action: museum_hours("City Museum"). After the observation arrives: “In the official hours record dated 1 September, Monday is shown as closed.”

What did we get?

The decision did not rest on an observation invented before the tool call.

Medium

Situation

The first document has no application date; a second source becomes necessary.

Prompt

Task: Find the application deadline for the fictional exhibition.
The supervisor should run at most 3 read-only document-reading calls. Each round should carry the task, the IDs of the documents read and the quotes to the model.
1. read_document("exhibition-notice") -> "For the date, see the conditions document: exhibition-conditions."
Ask the model for a single next action; do not return to the same document without reason.
2. read_document("exhibition-conditions") -> "Applications close on 20 October 2026 at 17:00."
Model: Write the result with the document ID. If the dates conflict, do not give a definite date; if the budget runs out, report the gap.

Sample output

The second action is read_document("exhibition-conditions"). Result: “20 October 2026, 17:00 — exhibition-conditions.”

What did we get?

The first observation determined the second step.

Hard

Situation

A support diagnosis must tell successful and failed queries apart.

Prompt

Task: Why is order K42 not showing? There is no permission to write or resend.
Supervisor: At most 3 tool calls. Permitted tools order_read(id), sync_status(id). State: call ID, response status, finding, remaining budget.
order_read("K42") -> {status:"not_found"}
Model: Do not interpret a missing record as a deleted order. Choose the next read-only check.
sync_status("K42") -> {status:"pending", next_retry:"14:30"}
Model: Evaluate this result together with the earlier observation; write the limit of the diagnosis and the follow-up time for the customer. Do not rerun the tool; stop once you have produced a status explanation.

Sample output

“No record was found in the order lookup; it is waiting in the sync queue. The next automatic retry is at 14:30. There is no evidence that it was deleted.”

What did we get?

The failed query and the actual cause were not confused with each other.

Where should you stop?

ReAct does not create tool access. The supervisor must enforce permissions and the call budget. Unlike a fixed-step prompt chain, the next action is chosen based on the observation. Instead of asking for a dump of internal thinking, a record of decision, action, source and result is enough.

Sources

• ReAct: Synergizing Reasoning and Acting in Language Models (new tab) — Yao, Shunyu; Zhao, Jeffrey; Yu, Dian; Du, Nan; Shafran, Izhak; Narasimhan, Karthik; Cao, Yuan. 2022-10-06; version read 2023-03-10. Defines the ReAct mechanism in which decisions, actions and external observations alternate; the examples are simple adaptations written for this collection. Evidence level: relevant body sections of the original paper.

  Copy source address