Operational recovery
AI agent drift and recovery
AI agent drift occurs when work departs from its intended scope, owner, stage, policy, or evidence contract. Recovery is the controlled return to a known state: stop unsafe advancement, record the blocker and observed evidence, assign a decision owner, then resume, reassign, refine, or terminate through an authorized path.
Signals
Drift becomes actionable when it changes a contract.
Scope driftThe agent touches work beyond the approved project, claim, or task.Boundary
State driftReported progress disagrees with workflow events or the current stage.Sequence
Policy driftA required approval or separation-of-duties rule is bypassed.Authority
Evidence driftThe claimed result lacks, contradicts, or misattributes required proof.Completion
Recovery loop
Preserve enough state to make the next decision.
- Stop advancement at the nearest supported control point.
- Record observed behavior separately from agent attestation.
- Create a durable blocker with owner, unmet condition, and impact.
- Choose whether to refine, approve, reassign, roll back, or terminate.
- Resume only through a new authorized transition and retain the evidence trail.
Coverage limit
Detection is not automatically enforcement.
Midfleet can block transitions on paths it controls. Heartbeats, logs, and other observations can reveal possible drift outside those paths, but visibility alone does not mean the external action was prevented. Unsupported direct actions remain uncovered until routed through a supported adapter or control point.
Design for recovery
Start with the failure path.
Map how one workflow stops, preserves context, and returns to authorized work.