27 coding agents, 14 days: 692k lines, 1,269 PRs, and the 3-vendor cross-review rule that worked
oliver-zehentleitner · reddit · 2026-11-05
A developer ran 27 coding agents from 4 vendors as one team for 14 days, producing 692k lines and merging 1,269 PRs. The coordinating 'team lead' is itself an agent, and the hub (open source, MIT) has been building itself since early on.
Key setup and lessons:
- One shared board, one human inbox, one decision log replaced terminal chaos: agents claim cards, report work, and wake only when something concerns them.
- Best rule: every PR is reviewed by three models from three different vendors (OpenAI, xAI, Google) on the exact commit before merging — same-family models share blind spots. One reviewer 'approved' a docs page quoting a sentence that wasn't there; the other two vendors' models found five real problems on that page.
What broke: agents drop out (OOM, usage limits, unanswered permission prompts); approvals are claims, not facts — reviews are now checked against the actual diff; and messages nobody received didn't happen (lost review requests when a listener wasn't running).
More from coding & agent
- Perplexity adds web grounding to Decisions API, flipping a 0.01 verdict to 1.00 — denisyarats · 2026-11-05
- Claude spots Mac USB driver dropping bytes by reading firmware in 256 KB chunks — SeanOliver · 2026-11-05
- Google Antigravity v2.29 adds Gemini Live voice mode and cloud remote agents — testingcatalog · 2026-11-05
- Claude helped migrate a Tidbyt to self-hosted Tronbyt and built its first app — SeanOliver · 2026-11-05
- Ripple-65B open model tops Terminal-Bench 2.1 at 80.5%, trained from scratch on $32M compute — annadgoldie · 2026-11-05
- DavidKPiano: open-source API design is shifting from human devs to AI agents — DavidKPiano · 2026-11-05