Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresMean final score of a team of 2 agents in the same four hours, which the report compares with the team of eight to measure what more agents and coordination add; a diagnostic. Bespoke Labs' Multi-Agent Coordination Bench (November 18, 2026) gives agent teams twelve natively multi-agent, long-horizon tasks with no assigned roles: Mars Return, 0 A.D., Holdout-game, Mindustry, Power Outage, Seed to Profit, Cave Rescue, Dubai ATC, Chicago Skyline Survey, Spectrum to Molecule, RNA Secondary Structure and SOTA Compressor, nine on a shared repository and three distributed. Every run lasts four hours of wall-clock time whatever the team size, agents coordinate through a message board, and a hidden deterministic grader scores the result between 0 and 1 against fixed anchors (0 A.D. has no stated cap). The score is the mean over the tasks of each task's mean final score, usually over three runs, with each model in its native harness.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | max · Claude Code | 0.4 | |
| 2 | Gemini 4 Argon | high · Antigravity CLI | 0.3 |