Agent Budget Overrun Detector

Scenario

Use after a ledger is available and before a loop, retry policy, or multi-agent fanout is allowed into a wider rollout.

Input and output

Input contains redacted run summaries with cost, retry count, loop count, context units, and thresholds. Output has risk codes, per-run findings, and a remediation queue.

Execution and recovery

Run the README command. Missing thresholds or malformed rows return a visible error. A finding is evidence for a human to change the workflow, not an automatic kill command.

Privacy boundary and acceptance

Keep only numeric summaries and synthetic identifiers. The sample reports a loop, retry, cost, and context-growth finding. It never stops a real run or contacts a provider.