ReAct
ReAct: Synergizing Reasoning and Acting in Language Models
Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao
The paper in one sentence
ReAct turns tool use into a feedback loop: reason, act, observe, and revise.
What the paper does
ReAct prompts language models to interleave reasoning traces with actions taken in an external environment. Observations from tools then feed the next reasoning step. The paper evaluates this pattern on question answering and interactive decision-making tasks.
Why it matters to engineers
Many agent systems still use this control-loop shape. The engineering challenge is not merely prompting a model to call tools; it is constraining actions, preserving state, handling failures, and making the trace inspectable.
Key claims from the paper
- Reasoning traces can help a model plan and update its approach while actions obtain external information.
- External observations can reduce reliance on unsupported internal knowledge.
- Interleaving reasoning and acting improves interpretability because the task trajectory can be inspected.
Architecture, step by step
- 1
Task state
The model receives the user goal, available actions, and the current interaction history.
- 2
Reason
The model forms an intermediate plan or decides what information is missing.
- 3
Act
A constrained action invokes a search tool, environment operation, or domain API.
- 4
Observe and repeat
The returned observation becomes input to the next decision until the model answers or stops.
Production takeaways
- Treat tools as typed capabilities with validation, authorization, timeouts, and idempotency rules.
- Separate model-proposed actions from policy enforcement and execution.
- Set explicit step, cost, and time budgets to prevent unbounded loops.
- Log action inputs, observations, and final evidence so failures can be reproduced.
Limitations and caveats
- Reasoning traces can be wrong, verbose, or unfaithful to the mechanism behind the final answer.
- Tool errors and adversarial observations can compound across multiple steps.
- The paper’s benchmark results do not by themselves establish safe autonomous operation in open environments.
Interview questions to test understanding
- Where should authorization be enforced in an agent tool loop?
- How would you detect and stop an agent that is no longer making progress?
- What is the difference between a tool error and a model-planning error?
- Which parts of an agent trace should be retained for evaluation and audit?