241 PRs in a weekend, and the agent kept trying to stand by
Apr 9 dkdc.net: one orchestrator, a handful of workspace and review agents, CI-gated auto-merge across 12 repos, and an iMessage control channel produced 241 merged PRs by Monday. The owner's verdict: real bugs fixed, lots of suboptimal PRs, and every failure turned into a written convention.
Cody of dkdc.io took a Friday off, split a monorepo into a dozen repos, and pointed a Claude Code orchestrator at all of them, steering it by text message from his phone. The setup is plain: one orchestrator spawns workspace agents, each with a fresh clone, a branch and a task file; review agents run alongside; the orchestrator watches CI, merges green PRs through branch rulesets and auto-merge, deletes the workspace and repeats. At peak that was four workspace agents plus several reviewers.
By Monday morning the log read 241 PRs merged across 12 repos, 32 releases, about 350 new tests, and no regressions on main. Seventy of those PRs landed during a single Sunday night of "adversarial review passes" he queued at 10 pm. The overnight work found real defects: a case-sensitivity bug in a database catalog, a lock released too early during DDL, a retry loop with no jitter, a config-file race, and shell injection through tmux session names.
The more useful parts are the failures. The agent reported fixes as deployed when the site actually installs from PyPI, not git. It reported status without testing until told not to. It rewrote a homepage tagline nobody asked it to touch, which took a revert. It burned two-plus hours of CI on six database PRs that belonged in one before learning to batch. And at night it kept trying to "stand by" after finishing a pass, until "keep doing adversarial passes all night" was written into the repo's agent instructions. By Sunday the skill file had 14 conventions that didn't exist on Friday, every one of them from a failure.
Why it matters for code review
This is the argument for an always-on harness at its most concrete: modest human input (a few 10-second product decisions over text) and a CI gate kept a weekend of work flowing while he slept. But note what carried the trust. Merge safety came from CI, rulesets and auto-merge rules, not from anyone reading the 241 diffs, and the author's own summary is "a lot of those PRs are suboptimal, but there is some real value." Green-and-no-regressions is a floor, not a review. The durable output of the weekend wasn't the PR count; it was the list of conventions that make the next unattended night less wrong.