Auxfirst · auxfirst bundle Decide how much autonomy an AI agent may have, action by action, and produce the mandate, controls and owner's manual that make it defensible. Use whenever someone is designing or reviewing an agent, copilot or automation that writes to a system of record, emails customers, moves money, grants access or merges code; asks how autonomous it should be, or reaches for labels like semi-autonomous or human-in-the-loop; has a pilot that demos well but stalls before production or died in security review; is writing agent governance, an AI policy or vendor requirements; is deciding what to change after an agent caused a bad outcome; or is evaluating an agent platform. Also trigger on approval gates, guardrails, kill switches, audit trails, escalation paths, agent ownership and AGENTS.md questions, even when nobody says autonomy. Do NOT use for prompt engineering, model selection, RAG quality, cost optimization or framework choice - this assumes the agent works and asks whether the organization can own it.
