ParallACT: parallel coding agents fail more on conflicting tasks
The point
Separate requests to coding agents in a shared repository can pass alone but conflict when merged. ParallACT tests this with no orchestrator, using complementary tasks that must share an interface and contradictory tasks with incompatible edits.
Why it matters
For builders, parallel agents in one repo need coordination beyond a channel.
Key details
- Contradictory tasks: Claude Code with Claude Opus 4.8 fell from 62 percent to 16 percent
- Same gap in six configurations pairing two harnesses with three open weight models
- A communication channel recovered only a small part at extra token cost
Original ~8 min → 25 sec here Source: arXiv cs.AI, Nov 3, 2026