Hellström: an overnight fleet merged 11 PRs, sat idle 94.5% of the night, and had no quality gates

May 6 from Felix Hellström: the first unsupervised night of his agent orchestrator ran 686 cycles in 7.5 hours, dispatched 103 tasks against 54 issues, auto-merged 11 PRs, flagged 9 for review, and killed and re-dispatched 8 stalled agents. Every failure lived in the harness, not the agents — and auto-merge on a green suite was luck, not a guardrail.

Felix Hellström let his orchestrator run unattended for the first time on the night of March 1, 2026, and wrote up the numbers. Over 7.5 hours the loop ran 686 orchestration cycles, pulled 54 issues from a backlog, dispatched 103 tasks, and auto-merged 11 PRs: dependency upgrades, bug fixes, feature work and new tests, each passing the suite before merge. Nine more were flagged for manual review. Eight agents stalled mid-task and were detected, killed and re-dispatched without him touching the keyboard.

The number he puts beside the 11 is 94.5% idle. The fleet did real work in 5.5% of its runtime and spent the rest polling an empty queue.

The setup is three layers: a scheduler that prioritizes backlog issues by label and assigns them to agent slots; Claude Code instances, each in its own tmux pane with one issue and a fresh checkout; and a supervisor that treats a pane with no output as stalled. He trusted that last layer least, because it reads a side effect (terminal output) instead of a real status signal, and it was the first thing he rebuilt.

What broke, with nobody awake to paper over it:

No adaptive polling. The loop kept checking a drained queue on a fixed interval. Back off when empty, ramp up when work arrives.
No wall-clock timeout. One agent ran for more than six hours on a single task and burned most of the night. His fix: kill and re-dispatch anything past 30 minutes on one issue.
Duplicate dispatches. 103 dispatches for 54 issues. Two cycles inside the same polling window could grab the same issue; issues have to be locked before dispatch, not after.
No quality gates. PRs auto-merged on a green suite with no diff-size ceiling, no post-merge CI, and no human gate above a complexity threshold. "For a first run I got lucky. Luck is not a guardrail."
Destructive restarts. Reconciling state on startup killed every running agent and lost in-flight work.

None of those failures is about the agents writing bad code. Every one lives in the infrastructure around them: scheduler, dispatch lock, health check, reconciliation. The morning after, the run's own failure log became the next backlog.

Why it matters for code review

This is the honest ledger of an always-on factory's first night: real merges while the operator sleeps, plus a harness that wasted most of its clock, double-booked work, and merged on green tests alone. The compounding comes from fixing the loop, and the merge gate is the piece that can't stay on luck. Compare Saagar Patel's night fleet and its verification bill, Relux on goal mode burning tokens while idle, and Winslow's point that coordinator-done is not merge.