Definition
Plain language
A test that scores an AI purely on how well it manages a team of helper agents — handing out tasks and access — rather than on solving the problem itself.
As stated in the literature
A benchmark isolating an LLM's subagent-management skill by fixing the worker pool and scoring task correctness times management quality, including least-privilege file and tool scoping checked against a real filesystem.
Why it matters: It matters because coordinating and safely scoping a team of agents is a distinct skill from problem-solving, and this measures it directly.
For example, it checks whether an AI manager gives each helper agent only the files it needs for its task and no access beyond that.
Heard on the show
“The paper is called ClawArena-Team, out of UNC Chapel Hill and collaborators, posted at the end of June 2026.”Episode 190 — The Skill Every AI Manager Is Missing: Handing Out Exactly the Right Keys