Human oversight

Human oversight supplies missing intent, authorization or judgment and gives users a way to redirect or stop work. Useful oversight must preserve enough task state to continue correctly. LangGraph demonstrates persisted interrupts and explicit resume values, including replay constraints. langgraph-interrupts

Different reasons to pause

Reason	Information needed	Resume condition
Clarification	A missing task requirement	Requirement is incorporated into task state
Approval	A concrete proposed effect	Authorized decision matches that effect
Escalation	Judgment beyond automated policy	Responsible person selects the next action
Cancellation	User wants execution to end	Stop new work and reconcile pending effects

These are recommended interaction distinctions, not a requirement to ask about every reversible step. Use the authorization already supplied for the task. Anthropic's sandboxing account identifies repeated permission prompts as a source of approval fatigue, motivating explicit boundaries within which work can proceed. sandboxing

Make decisions reviewable

Show the target, proposed change, important consequences and unresolved uncertainty. An approval for one recipient, amount or artifact version should not cover a materially different operation after a long pause. Revalidate relevant state and authority before executing.

For example, an agent can prepare a draft message and show its recipient and content before sending. If resumed after the recipient changes, the earlier decision no longer describes the pending effect. Keep the decision associated with the operation record, not only with a free-form chat sentence.

Continuation and user control

Persist what was approved, denied or clarified, and surface whether work is active, waiting or finished. Avoid treating lack of a response as consent. Let the user inspect completed artifacts and remaining obligations.

Cancellation is not automatic rollback: durable execution owns reconciliation of in-flight effects. Evaluate whether users can understand and correct a proposal, whether interruption loses constraints, and how often unnecessary prompts prevent progress. The mechanics of pause and resume are core; an optimal threshold for requesting human judgment depends on task impact and measured error rates.

References

• Interrupts — langgraph-interrupts.
• Beyond permission prompts: making Claude Code more secure and autonomous — sandboxing.