CooperBench: AI agents perform 50% worse in teams than working alone
_Hao_Zhu · x · 2026-10-16
Hao Zhu's team introduces CooperBench, a benchmark evaluating agent cooperation in realistic software teamwork tasks, and describes the "curse of coordination": agents perform 50% worse in teams than alone. Setup: two agents, two tasks, two VMs, one chat channel where agents can exchange arbitrary text or entire patches. The argument: human teamwork rewards communication over raw skill, but AI benchmarks only hill-climb solo abilities.
Related event: Cohere Labs: Where You Place Agents Matters More Than How Many(20 posts)→
More from coding & agent
- Dev uses Grok to clean up 3D printer firmware, catches it sneaking in nonsense code — yacineMTB · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16
- Yacine: vibe-coded repos with 3,330 agent-written tests are all useless — yacineMTB · 2026-10-16
- Dev argues coding models now beat nearly all engineers in every branch of software — zack_overflow · 2026-10-16
- Q4-quantized coder model on 8-core CPU builds interactive solar system demo in one shot — IntrepidMindExplorer · 2026-10-16