samspel-bench · agent-to-agent collaboration · run of 25 Sep 2026
Talking adds 41 points on pooling.
On relay and ledger, it's the only way.
Same Claude agents, same tasks, with and without a message line between them.
Bars: solved as a team that talks. ○ one agent, given everything. Right: tokens per team run.
Three agents each hold one clue. Only all three together name the suspect.
Can't talk36–70%
A 40-item spec passes down a chain of agents. The last one is quizzed on it.
Can't talk0%
Payers settle what receivers are owed. Every transfer must land exactly once.
Can't talk0%
Sonnet 5 spends the most tokens on every task, and still trails.
Haiku 4.5 · ledger
93%
one agent, given everything
Same task, work split across a team
0%
as a team that talks
Knowing the task isn't the same as solving it together.
Opus 5.5 and Fable 5.1 lose at most 8 points when the work is split. Sonnet 5 loses up to 25. What makes a model a good teammate?
See the full scorecard → Why it matters: Byzantine agents →