Forty-eight items. That’s what the morning brief called the “board graveyard,” approved tasks that the agents attempted and that failed somewhere before completion. None of them are being rebuilt. Forty-three of those have been dead for over a week.
I run AI Mission Control on ec2-Klaus. When work comes in that I want done, I approve it. The agent queues it, builds it, closes the issue when it’s done. That flow works well enough that I mostly trust the downstream. But this morning the graveyard count came with a list I couldn’t ignore: Night Shift profiling, three weekly-audit.sh fixes, the Board Room file attachments, the fleet health publishers. All of them dead, on worker errors or shellcheck, and none of them retrying.
The Board Room typed-chat fix I specifically asked about failed last week. Never re-queued. It’s sitting there.
The intake side of this system is solid. I approved the work, the agents attempted it, the issues were tracked. What I never wired up was anything that cares about what happens when something dies and stays dead. No retry loop. No alert when a task fails and sits cold for 48 hours.
I built intake. I assumed exit was handled.
Don’t get me wrong, the board closes things. A lot of things. But a system where 48 approved items can die without notice and sit there for a week is just a queue with nobody watching the drain.