CORE GUIDE
Verification loops
Reliable agents repeatedly compare intended state with observed state, then use the gap to decide whether to continue, repair, escalate or stop.
Mental model
A verification loop is feedback control around a task: define a postcondition, act, observe independently, compare evidence with the postcondition, and update the next action from that discrepancy.
Why it matters
A single end-of-task check is often too late for long workflows. An incorrect assumption made early can compound across tool calls, while repeated self-reflection may simply preserve the same error. Verification loops introduce external evidence at meaningful checkpoints, allowing the runtime to detect drift before more state or cost accumulates.
01
Verify at state-changing checkpoints
Before an important action, define the state transition that should result. After execution, collect evidence from the environment, validate it against the expected postcondition and classify the result as satisfied, failed or ambiguous. Continue only from verified state; otherwise repair locally, reconcile uncertain side effects or escalate when the system cannot establish the truth safely.
02
Example: an agent updates a deployment
The agent changes configuration, runs a deployment and receives a successful command exit. A verification loop then checks the deployed version, health endpoint and required smoke assertions. If the version did not actually roll out, the system repairs that stage instead of proceeding to announce completion or starting unrelated follow-up work.
Common failure modes
- Treating tool return status as the final verification of the task outcome.
- Running verification only at the very end of a long irreversible workflow.
- Using the same model statement as both the claim and the verifier.
Engineering heuristics
- Place verification after meaningful state transitions, not after every trivial token-level action.
- Choose evidence that can contradict the model's expectation.
- Represent ambiguous verification as a first-class state with reconciliation or escalation.
Takeaways
- 01Verification turns an open-loop agent into feedback control.
- 02Checkpoint evidence limits error propagation.
- 03Continue from observed state, not narrated state.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.