My agents broke each other's work 3 times in 7 weeks — four ownership rules fixed it
Humaux · reddit · 2026-11-09
A dev coordinating a team of coding agents recounts three failures in seven weeks, all from shared state with no ownership rules, not bad code:
- A subagent overwrote the whole shared task board, wiping four lanes of progress
- A reviewer flagged a parallel card's legitimate output as unauthorized because it only knew its own card's whitelist
- A DB check using (jsonb->>'key') IN (...) never failed since a missing key returns UNKNOWN, letting empty records pass
The fix was four rules, not smarter agents:
- One writer per element, by id; whole-board updates banned
- Reviewers judge only their own card's diff and direct callers
- A check that can't fail is not a check — verify it fails against a broken copy before trusting it
- Missing data is "unverifiable," never "pass"
Key line: most multi-agent setups aren't teams, they're five interns sharing one whiteboard with no names on it — with shared state, the ownership rules are the product.
More from coding & agent
- Sampling ≠ Intelligence: agent PRs merge at 32.7% vs 84.4% for humans, argues dev — gerardsans · 2026-11-09
- First AI opera adaptation of Frankenstein built on a coding-driven generative pipeline — hq4ai · 2026-11-09
- Engineer proposes cross-engineer agent-to-agent communication as human becomes the bottleneck — thomasahle · 2026-11-09
- Matt Pocock defends 'Software Factory': top teams now build systems that build the product — mattpocockuk · 2026-11-09
- Day 309/365 of GPU Programming: Optimizing LLaMA2-7B on Edge with SIMD and Multithreading — NandoDF · 2026-11-09
- Berkeley-led paper: evolved agents gain 5.2 reliability points while losing 10 tasks they once solved — rohanpaul_ai · 2026-11-09