Observations help revise the next action

The important loop is not simply writing a longer plan. It is using external feedback to revise the next step.

An original search assistant could inspect a source, recognize that evidence is missing, and search again with a narrower question.

External feedback still needs trust boundaries

An observation can be incorrect or malicious. Keep source content separate from authorization, and verify that a proposed action remains within the user’s task.

What the paper showed

On question-answering and interactive tasks they built, letting the model alternate between a reasoning trace and a tool call beat several baselines that only reasoned or only acted.

The trace is supposed to help the model decide the next action and to make the run readable. The observation comes from the tool, not from the trace.

If the tool is wrong, the trace will often explain the wrong fact with confidence.

The pattern is now ordinary: agents that call search, code, or browsers. Ordinary does not mean the paper’s gains transfer.

Their tools were simple and their tasks were curated. A production tool with side effects is a different risk.

The trace can also leak into the user interface as if it were an audit log. It is a generated monologue.

The audit log is the tool call record.

What to copy

Copy the discipline of writing the next action separately from the prose, and of storing the observation verbatim. Do not copy a claim of reliability.

Add the checks their environments did not need: permission, idempotency, and a stop when the same action repeats. Cite Yao and colleagues for the pattern, and your logs for the behavior.

What to record before you rely on it

If you say a system is ReAct-style, show a trace with the action and the observation marked. Keep the tool log as the record.

Do not cite the paper as the system’s success rate.

Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.

When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.

Name a person who owns that failure.

Common questions

Is chain-of-thought the same thing?

Chain-of-thought has no tool. ReAct adds an action and an observation from outside the model.

The outside part is the difference. Without it, you have a monologue.

Does the reasoning trace justify the action?

It explains what the model said it was doing. The justification, if any, is whether the observation supports the final answer.

Read the observation.

Sources & further reading

  1. ReAct: Synergizing Reasoning and Acting in Language Models ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories