Problem it solves

When an agent pauses due to an error, users see a failure state with no information about what happened before the pause or what they can do next.

When to use

Whenever an agent execution halts unexpectedly — due to schema mismatch, policy violation, unresolvable step, or external system failure.

When not to use

For planned pauses requested by the user or for approval gates. Those are not errors.

Governing principle

A pause is not a dead end. The recovery surface must show what completed before the pause, what the agent was attempting, and what the human can do — with all three choices available: continue, rollback, escalate.

Required Components

Interaction Flow

1

Agent pauses unexpectedly

Execution halts. The agent state transitions from RUNNING to PAUSED.

2

Attention trigger surfaces

An Agent Attention Trigger immediately notifies the user of the pause and its cause.

3

Recovery surface opens

The Recovery & Override component shows what completed before the pause, what the agent was attempting, and the specific error.

4

Human reviews options

Override and continue (if the error is non-critical), rollback to last stable state, or escalate to senior review.

5

Decision executed and logged

The chosen action is executed and logged to the audit trail with timestamp and user identity.

Governance requirements

Every agent pause must be logged with the step that failed, the error type, and the subsequent human decision. Rollback events must record the state restored to.

Accessibility notes

Pause states must use role="alert". Recovery controls must be keyboard navigable. The active choice must be visually distinguishable through more than color.