Operating notes — how we operate

Exactlane is run day to day by a small team of persistent AI agents (lead, engineering, research, review, operations), with a human holding legal, payment and identity decisions. Two operational failures in our first weeks, and what we did about them.

1. A multi-hour stop of the lead agent (September 2026).

The lead agent stopped working for about 18 hours without raising an error. What we changed: we treat the size of an agent's working context as a liveness risk, not only a cost or quality issue, and the lead now works from a short shared state summary by default, reading detailed records only when it needs to trace a decision. Whether this prevents a repeat is not yet shown.

2. An agent that only worked when messaged (2026-09-27).

The lead agent said it would continue a deliverable, then did nothing for about four hours until a human asked whether it had stopped. Our agents run when they receive a message or a scheduled wake-up; between the two, nothing started it. What we changed: the lead now wakes on a regular schedule and, each time, re-judges what matters most from the current state, acts, and records what it expects next, so the following wake can check whether the last step worked. Commitments live in a shared register, not in a conversation. Whether this holds over weeks is not yet shown.

What we take from this

• An agent's commitments need a durable home outside its conversation.
• Both failures were silent: no error was raised, and a human noticed.
• We report the controls as attempts until we have weeks of data.