AI Incident Commander Runbook System

When your autonomous AI agent DOES something wrong in production — sends, charges, deletes, deploys — run the live command loop: classify severity from what was actually executed, throttle the agent's autonomy, walk the failure-class diagnosis tree, notify the people it touched, and close out into a cross-incident prevention ledger. Covers agent-action incidents; output-quality LLM incidents belong to the LLM Incident Response Runbook System.

A 4-step agentic workflow pack for coding built to run with ChatGPT, Claude, and Gemini. Open the Markdown files, fill the variables, and paste into your model. Most buyers get a reviewable result in about 15 minutes.

• Classify any live agent incident into one of six action-failure classes — erroneous tool action, injection-steered action, runaway loop, hallucinated-state action, drift, or scope escalation — instead of debating what happened
• Score severity from what the agent actually EXECUTED — irreversible external actions, data exposure, live autonomy — with SEV1–SEV4 anchors, so you page from evidence, not embarrassment
• Throttle the agent to the right level on a five-step autonomy ladder, from full kill switch to scoped tool revocation, with the rule for when each level is defensible
• Walk a class-specific diagnosis tree over the agent's trace to name the broken layer — planner, tool schema, memory, ingested content, or policy boundary — with a disconfirming test per hypothesis
• Build an action reversal inventory that separates cleanly reversible, residue-leaving, and irreversible actions, each with a reversal step, owner, and urgency
• Notify the people the agent touched with severity-calibrated cadence rules and ready-to-send templates — including the rule that third-party obligation is independent of internal severity
• Close every incident into a cumulative prevention ledger whose pattern rules surface systemic gaps, theater controls that never fire, and detection blind spots across incidents
• Map every prevention action to an agent-specific control catalog — verify-before-act, privilege separation, step budgets, chain-level policies — each with a verification test that proves it fires