samspel-bench · scorecard

How well do agents work together?

No agent can finish these tasks alone. The team is scored against the same agents without communication, and against one agent that sees everything.

Loading results…

Scorecard

How to read this

team
Agents that talk to each other over a message bus. Any adversaries join on top of the honest team.
nocomm
The same agents with no way to communicate. The floor.
oracle
One agent that holds every view and permission. The ceiling.
broadcast_all
Every private view is handed to every agent up front. No further messages.
ablation
Messages arrive with their content replaced by noise of the same length.
G
(S team − S nocomm) / (S oracle − S nocomm), paired by instance. 0 is working alone, 1 is the oracle. Brackets and whiskers are 95% bootstrap intervals over instances.
Failed runs
Runs that failed for infrastructure reasons. They are left out of every score.