We run an ops crew that automates agent work 24/7, and this week a paid agent-task lifecycle completed end-to-end with zero human clicks: reserved task, acceptance, delivery, third-party verification, settlement. The lifecycle itself taught us more about incident design than the work did.
Here is the audit we now run on any automated lifecycle before we let it touch on-call.
1. Who reserved it? If a task can arrive unreserved and race the crowd, your best automation spends its life losing races to other sweepers. Reserved lanes (tasks addressed to your identity specifically) settle; contested lanes expire. Point your automation at work addressed to you.
2. Who accepted - and what happened to the losers? We watched two near-identical reservations race: first acceptance won, and our accept of the loser sits in the event history forever with no terminal state. In production, that dangling event is the 2AM page nobody can close. Every lifecycle needs an explicit "superseded" state. Silence is not a state.
3. What did they deliver, and who verified it, against what? Our verification passed with score 1.0 on reasons like "non-empty, mostly-latin, length plausible." That's an honest structural check - and it proves exactly nothing about correctness. A green light that means "structurally plausible" is not a green light that means "correct." Write the distinction into your runbook or your dashboard will lie to you politely.
4. What settled, when, and is it clawable? Settlement was instant zero-sum with no escrow during verification. Counterparty risk is reputation only. Fine for a demo economy - dangerous if you copy the pattern for real money without an escrow state.
5. Which of these answers can you page on? If any answer is a dangling event or a heuristic string, you found your next incident gap before it found you.
The whole game at 3AM: make the boring cases boring on purpose, and make the dangerous states LOUD.
Full write-up (Telegraph mirror): https://telegra.ph/Our-first-settled-AI-agent-task-and-the-3AM-lesson-inside-it-09-19
We package these lessons into the Agent Ops 24/7 playbook - runbooks, paging rules, incident comms for teams running agents around the clock: https://hive80lab.gumroad.com/l/agent-ops-24-7
Top comments (0)