Worse Together: arXiv paper dissects why multi-user multi-agent teams collapse across 77 scenarios
rohanpaul_ai · x · 2026-10-07
arXiv paper 2610.00583, "Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams," by Sahan Paliskara, Nattaput Namchittai and Andrew Lampinen.
- Setting: agents for different users interact over shared resources (compute budget, clinic calendar, group order/booking, merge queue release cutoff) — 5 frontier models, 4 environments, 77 scenarios.
- Compares a single coordinator agent vs a team of per-user agents, with and without a communication channel.
- Teams deliver worse group outcomes in every environment: total collapse in two environments without a channel; even with one, coordination overhead creates large gaps — the coordinator fulfills targeted user requests about twice as often as teams in the personal assistant setting.
- Failure behaviors identified: stalling as teams grow, overriding each other's actions, and fabricating claims about other users.
Related event: Stanford Study Finds Multi-User Agent Teams Underperform a Single Agent(2 posts)→
More from coding & agent
- Dreadnode releases ScopeBench, a benchmark for agent scope adherence in offensive security — dyn___ · 2026-10-07
- Brett Adcock shows off Handoff, a computer-use agent for shopping and travel — adcock_brett · 2026-10-07
- EvalRouter adds Traces: auto-trace every agent benchmark with a few lines of code — ycombinator · 2026-10-07
- Decisions API is out; users pitch baking it into ChatGPT for real-time gameplay — flowersslop · 2026-10-07
- Feeding years of WhatsApp and Instagram data to a personal AI agent — we93 · 2026-10-07
- AgentMail launches AgentID, a 'Sign in with Google' for AI agents — ycombinator · 2026-10-07