ReAct: decisions grounded by tool observations

Yao et al.'s ReAct paper interleaves generated reasoning, actions, and observations in question answering and interactive environments. Treat its trace as a control pattern: model text proposes the next action, while a tool result supplies external evidence. It does not prove that generated rationale is faithful, that a tool call is authorized, or that grounding guarantees correctness.

Use short decision summaries: goal, uncertainty, chosen action, and what observation would resolve it. Do not require a model to expose private chain-of-thought; retain the tool request, tool result, citations, and a concise user-safe explanation sufficient to audit a result.

Evidence-carrying loop

flowchart TB
  G[User goal] --> D[Decision summary: gap and next action]
  D --> A{Action authorized and valid?}
  A -- no --> Z[State limitation or abstain]
  A -- yes --> O[External observation]
  O --> E[Evidence ledger: source, time, limits]
  E --> U[Updated decision summary]
  U --> C{Claim supported?}
  C -- yes --> R[Answer with cited observation]
  C -- no --> Q{Justified follow-up within budget?}
  Q -- yes --> D
  Q -- no --> Z

sequenceDiagram
  participant U as User
  participant M as Agent
  participant T as Authorized tool
  participant L as Evidence ledger
  U->>M: factual question
  M->>T: scoped query
  T-->>M: result plus source
  M->>L: record request, result, source, limitation
  M-->>U: claim tied to source or abstention

Compact trace fixture

goal: identify whether a claim is supported
decision: source A is incomplete; query source B for the missing fact
action: search({query: "...", scope: "approved corpus"})
observation: {source: "B", excerpt_id: "B-17", result: "conflicts with A"}
next: do not assert the claim; seek a primary source or say evidence conflicts

The decision is not evidence. The observation's source and retrieval context are. If a result is empty, malformed, or outside the authorized corpus, record that limitation and either reformulate a bounded query or abstain.

Design rules

• Validate action schemas and authorization outside the language model. Tool availability is not permission.
• Carry raw tool receipts or stable excerpts into the evidence ledger; a later summary must not be the only provenance.
• Use a state-derived loop guard (repeated action, unchanged observation, or budget selected by the application), then stop or escalate through an existing policy.
• Human correction may change a decision or final assertion; retain the correction and supporting evidence. It must not silently rewrite historical tool results.
• Benchmark figures in the paper apply to named datasets, prompts, and models; do not convert them to universal error or hallucination rates.

Source-grounded navigation

• Reasoning/action trace and limits
• Evidence versus generated text
• Failure propagation and recovery
• Uncertainty as a query or abstention trigger
• Interpretable operational traces
• Source evaluation and controller contract

Boundaries

ReAct is a prompting and interaction pattern, not a substitute for authorization, source validation, privacy review, or effect confirmation.

Bundle navigation

diagrams index.