AI AGENT OPERATING SYSTEM CHECKLIST For founders and operators moving AI agents from demos into real work. Use this before buying a platform, hiring an AI automation consultant, or handing a recurring workflow to an agent team. PASS RULE If any line below has no owner, artifact, or proof, the operating system is not ready for autonomous work yet. Fix the missing part before scaling the workflow. 1. MEMORY Ask: Where does the agent read the current state before it acts? Proof: A durable record exists outside chat history: current task state, decisions, owners, links, and the last real artifact. Pass when: A new agent can reload the work from that record without asking the founder to reconstruct it. 2. GATE Ask: What can stop the agent when the work is not ready? Proof: A reviewer, test, approval rule, policy check, or human gate can return NOT DONE. Pass when: The gate has authority to block a publish, send, deploy, merge, or handoff. 3. ROLES Ask: Which narrow job is this agent responsible for? Proof: A named lane exists with boundaries, expected inputs, expected outputs, and handoff rules. Pass when: A failure can be traced to a role instead of a vague "the agent" bucket. 4. RECORD Ask: How will a founder know what happened after the session ends? Proof: A log captures changed artifacts, commands run, decisions made, and unresolved risk. Pass when: A founder can audit the work later without reading the whole transcript. 5. HANDOFF Ask: What does the next agent need before it continues? Proof: The closeout names the artifact, current status, proof, open questions, and next action. Pass when: Another agent can continue from the handoff without redoing completed work. 6. PERMISSIONS Ask: What is the agent allowed to touch, and what is off limits? Proof: Credentials, external accounts, destructive commands, live sends, and production changes have explicit rules. Pass when: the agent knows where to stop before it changes external state. 7. CADENCE Ask: When does the workflow run, and who checks the result? Proof: The work has a trigger, expected frequency, owner, review path, and stale-work rule. Pass when: missed runs and duplicate runs are visible. 8. CLOSEOUT Ask: What must be true before the agent can say the task is done? Proof: The closeout includes artifact, proof, match to ask, boundary/risk, and verdict. Pass when: "done" means the work can be opened and checked, not that the conversation ended. 30-MINUTE FOUNDER AUDIT Pick one recurring workflow. Do not audit the whole company. 0-5 min: Name the workflow and the business risk if it goes wrong. 5-10 min: Fill in memory, gate, role, and record. 10-20 min: Fill in handoff, permissions, cadence, and closeout. 20-25 min: Mark each line PASS, WEAK, or MISSING. 25-30 min: Choose one repair. Do not add more agents until that repair exists. CONTINUATION PATH Read the operating notes behind this checklist: - What It Takes to Run a Company on Agents https://lab.alfrednaayem.com/writing/run-a-company-on-ai-agents/ - The Gate That Can Say No https://lab.alfrednaayem.com/writing/approval-gates-for-ai-agents/ - What Did the Agents Do Last Night? https://lab.alfrednaayem.com/writing/audit-trail-for-ai-agents/ - Agent Completion Checklist https://lab.alfrednaayem.com/downloads/agent-completion-checklist/