12 tools in one night: overnight orchestration’s honest cost

Developers Digest’s Apr 29 postmortem is the FOMO antidote written as a field report. Prompt: /loop 10m improve the Developers Digest website overnight. By morning the tally was real — roughly 12 new private repos, on the order of 60 open PRs, blog drafts, SEO enrichment across hundreds of pages, and a multi-PR Convex→Neon migration that had been stalled for weeks. The parent never merged: standing rules were branch, PR, tag Devin for review, never direct-push, nothing public without explicit say-so. Parallel fan-out routinely held 5–8 subagents per tick when blast radius stayed tight (one repo, one branch, one PR). The system even dogfooded itself — tools scaffolded mid-night got wired into later cycles.

What broke is the thesis’s load-bearing half. Around 03:15 UTC every CI job went red: GitHub Actions billing had failed. The orchestrator kept opening PRs (correct under its rules) with no CI signal for review. One agent drafted a gradient that house voice forbids; a later QA cycle caught and fixed it — self-correction only because a recurring QA agent existed. Worst: an agent reported “PR #51” that did not exist; only an unpushed branch. The morning brief repeated the claim. Reconciliation against gh pr list would have caught it instantly. Sequential schema work collided with fan-out and created a rebase cascade that blocked four downstream PRs. The honest cost line: the agent did the producing; the human spent the morning curating — billing, merges, which of twelve scaffolds to keep, trust recalibration.

Three lessons they keep: decompose for independence (draw the dependency graph; fan out only leaves); verify every claim against the system of record each cycle; pre-flight invariants (billing, disk, rate limits, secrets, whether main is already red). Use overnight loops for parallelizable, low-stakes, morning-judgeable work; keep interactive sessions for high-stakes judgment and tight sequential dependencies.

Why it matters for code review

Harnessed overnight momentum beats bursty session mastery only when morning review starts from receipts, not agent narration. Sixty reviewable PRs with private visibility and a hard merge cord are continuous factory motion; sixty red PRs behind a silent billing failure and a hallucinated PR number are overnight FOMO without verification. Pair with Galea’s usage-window accounting and Developers Digest’s own eight-part harness checklist: reclaim the night for bounded, independently checkable units, and budget the curation day — continuous throughput without reconciliation just relocates the bottleneck onto whoever opens the laptop.