The agent that never merges: overnight pipeline, human irreversible step
Sep 8 Allen Hutchison: scheduled auto-dev ticks triage→plan→build→address review; harness forbids merge — morning work is decide-to-ship, not reverse-engineer intent.
Source: The Agent That Never Merges
Allen Hutchison (Sep 8) describes the overnight factory two founders actually run at Vycari: wake to a queue of reviewed PRs, not an inbox. A scheduled Maintainerd auto-dev skill ticks a GitHub-label state machine — triage → plan comment → human plan approval → oldest-ready build → tests → open PR → work Greptile comments — and exits. Waiting happens between ticks. State lives on issue labels (auto:needs-info / planned / ready / …); no dashboard database. A phone-friendly review-queue skill collapses everything blocked on a human decision into one ordered list.
The load-bearing rule is mechanical, not prompt theater: the harness cannot merge. Two human gates — approve the plan before build, decide the PR is good enough to ship — keep irreversible accountability with a person. Hutchison’s failure mode is honest and useful: a burst of concurrent runs opened four PRs for the same bug before any claimed the issue; open-PR caps and reclaim rules held, race-to-claim did not. Cadence and concurrency are raised slowly so failures stay cheap. Public Gemini Scribe issues get the same pipeline with Claude-attributed comments and fresh VMs — awkward triage questions can land on reporters, but the never-merge rule caps blast radius.
Why it matters for code review
Steady harnessed overnight momentum beats peak-session heroics when “done” is evidence-gated and merge stays a human irreversible step. Continuous factories that auto-merge on green CI trade speed for a different risk; Hutchison’s design productively complicates FOMO — agents work 24/7 on your behalf, and you still own ship. Review the harness permissions and the morning decision queue, not whether Claude looked busy.