My AI agents had a bug that no test could catch: a job nobody's instructions knew about
I run my companies with a standing fleet of 13 AI agents and no employees. Each agent has a written contract, a schedule and a run log. They hand work to each other through queues, and every item they touch lands in one append-only ledger. In August the fleet went quiet in a way that looked healthy. Nothing crashed. Every agent ran on time and wrote its log. And for ten days, a job I had asked…
My AI agents functioned as a team without any human employees. Each agent had a contract, schedule, and log. They communicated through a queue and recorded each task in a ledger. In August, the fleet went silent, with a job I had requested never being completed. One agent was to draft a daily alert, while another, responsible for the morning summary, was supposed to present it.
However, the morning agent followed its own instructions and failed to show the alert. This oversight persisted for a week, with no one responsible for fixing the issue. The problem emerged from a handoff that existed only on paper and not within the agent's instructions. In such cases, agents refuse to act unless the task is explicitly stated in their instructions.
To address this, three rules were implemented: a handoff is only valid if the receiving agent has it in its instructions; new handoffs require editing the receiver's instructions, and all changes must be recorded. A switched-off agent with no successor is considered a defect. Every output requiring a human reader must be associated with its renderer, indicating who, where, and when the output will be presented.
A governance playbook ensures the map of responsibilities is updated nightly, checking for missing handoffs. This process adds a few minutes to each change but prevents similar issues. The details of the agents and their systems are available online.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
