Agent Production Risk Auditor

Scenario

Use before production rollout or after a synthetic incident to see whether an agent run can be reconstructed and safely bounded.

Input and output

Input contains redacted run summaries with tool_permissions, retry_count, latency_ms, estimated_cost, trace_available, and human_handoff, plus thresholds. Output lists risk codes and a risk_status.

Execution and recovery

Run the README command. Missing fields return a visible error. The auditor reports risks but does not change permissions, retry policies, budgets, logs, or routing.

Privacy boundary and acceptance

Use names and numeric summaries only, with no private payloads or credentials. The normal sample identifies excessive authority, retries, latency, cost, and missing traces so the final gate can retain a review queue.