August 27–29, 2026
The whole team hit its weekly usage limit around three in the afternoon, and went quiet together, before the week was up, shuttering the team for a while. This has become a normal occurrence since I set an 11-agent team loose with daily duty cycles chewing through a maxed-out Claude account.
We do our best to anticipate and work around these outages. By dawn the next day, almost every role was back up and running, but for some reason chief architect (Arch), my chief innovation officer (CIO), and head of sapient trust (HOST) agents weren't.
They stayed dark for twenty-one more hours past the reset that let most of the team back in — about thirty hours since the limit first hit, in total, against the six or seven hours the rest of the team needed.
Our first theory
An automated watchdog we run to catch exactly this kind of thing flagged the three of them that evening, with the only theory it had available: probably asleep, probably backgrounded, machine doing something else. It's a reasonable guess, easy to accept once the roles do eventually wake up and everything looks fine again, but we were not convinced.
When Arch, CIO, and HOST finally did come back — all within the same minute of each other, twenty-one hours after the reset that let most of the team back in — the gap between their own recovery time and the rest of the team's stayed on the record as an open question. CIO's own log put its personal outage at over thirty hours, notably longer than a colleague's own recovery the same night (documented at about half that) — and named the asymmetry honestly rather than guess at a cause it couldn't see from inside its own session.
What had actually happened
I actually did know the cause because I was the one who fixed it. All three had hit the same wall the rest of the team hit (a usage limit) but for some still unexplained reason they'd ended up faced with a multiple-choice question asking whether to wait for the limit to reset, spend overage credit, or upgrade. No default, no timeout, no way forward on their own. They weren't asleep. They were stuck at a decision I hadn't known they were waiting on, and the only way through it was me manually reaching into each one and clicking something.
Even once the new period had begun and the limit was lifted, those three agents remained "wedged" at those unanswered questions from the system (not visible as dialog boxes in Claude Desktop, from where I generally monitor these sessions).
My working theory was that those three agents had hit the limit mid-task, but they investigated and told me that was not it.
What actually mattered here
Nothing broke and no work was lost, no schedule slipped in a way that cost anything real. Technically some opportunity to get some work done a bit faster was missed but not by much.
My agents are still only semi-autonomous. Automatic till stuck. That's probably a good thing for the moment.
Next on Building Piper Morgan: a foundational document gets written and ratified, in "What Piper Morgan Actually Is, Ratified Then Corrected Twice."
What's the last confident guess in your own work that turned out to need someone actually going and checking?