Agent Drift & Self-Correction Auditor

The problem

AI agents can keep working long after the plan has stopped working. They repeat failed actions, preserve stale assumptions, ignore contradictory tool results, and sometimes change their explanation without changing the execution path.

What it does

• Re-anchors the agent to the actual objective and completion condition.
• Detects repeated failure loops, non-progress, stale assumptions, and tool-state mismatches.
• Separates what the agent knows from what it is merely assuming.
• Checks whether an error actually changed the plan or only changed the wording.
• Tracks who or what carries the cost if the agent continues incorrectly.
• Forces an explicit CONTINUE, VERIFY, RETRY, REVISE, ASK, ESCALATE, or STOP decision.
• Maintains a concise correction trace so the same failure is less likely to repeat during the workflow.

Why use this skill

Generic “reflect and try again” prompting can create the appearance of self-correction without changing the agent’s route. This skill makes correction operational: new evidence must change the next action, plan, confidence, verification requirement, route, or stop condition.

Use cases

• Long-running coding and research agents.
• Agents repeatedly calling failing tools.
• Autonomous workflows that wander away from the original task.
• Debugging agent loops and stalled execution.
• Adding explicit revision and stop gates to multi-step workflows.
• Reducing wasted tokens and repeated failed actions.

Known limitations

This skill cannot create persistent model learning by itself and does not guarantee correctness. Durable cross-session correction depends on the capabilities of the host agent, memory system, workflow, and available tools. High-stakes or destructive actions still require appropriate human review and authorization.