The subagent management bottleneck is granting privilege, not perception
Can a leader model manage a team when it only sees part of the workspace?
Opening
Production agents have quietly changed jobs. Instead of solving the task themselves, the main model creates specialized subagents, delegates work, and orchestrates their parallel, asynchronous returns through dynamic workflows — it is now a manager. Existing benchmarks score a policy’s own task-solving or a fixed multi-agent system’s emergent behavior, but none isolate the management ability of the single LLM acting as leader. That gap matters because this is where the failures land: wrong permissions, wrong routing, work handed to the wrong subagent. ClawArena-Team was built to measure the management skill itself.
What the paper does
ClawArena-Team is a benchmark of 41 multi-turn, multimodal, multi-directory scenarios spanning 258 evaluation rounds and 72 staged updates. The main agent is deliberately constrained: it natively perceives only text and directly accesses only part of the workspace, and it commands a fixed, locally served subagent pool, so score differences reflect management skill rather than raw capability. All scoring is execution-based with no LLM judge: the Subagent-Management Score (SMS) multiplies task correctness by a least-privilege and modality-routing factor. The authors run it across twelve proprietary, community-hosted, and self-hosted models and report where management actually breaks.
Findings
Management is the job being scored — The benchmark assumes the shape production has already taken: the leader does not solve the task, it staffs a team, assigns work, and collects the returns. That role, not the solving, is what the score rewards.
a main model creates specialized subagents, delegates work, and orchestrates their parallel, asynchronous returns through dynamic workflows
A fixed pool isolates the skill — With the subagent pool held constant and the workspace only partially accessible, any score gap between the twelve models is a gap in management, not in raw model capability. The design is what makes the comparison fair.
score differences reflect management skill, not raw capability
Privilege granting is the bottleneck — Perception is not what stops the leaders; deciding who gets access to what is. Across the study, no model exceeds 50% workspace-permission precision.
the management bottleneck is privilege granting rather than perception (no model exceeds 50% workspace-permission precision)
Caveat
The stated limitation is the setup itself: the main agent is deliberately constrained, perceiving only text and accessing only part of the workspace. The management scores therefore describe leadership under that restriction, and a leader with fuller perception could rank differently — the constraint is what makes the comparison clean, and also what bounds it.
What this means for 5ac
The orchestrator profile on the Kanban board is exactly the role ClawArena scores: it spawns worker profiles, grants them workspace access, and collects results. Two changes follow. First, permission precision is the weak link, so a card body should name the files a child profile may touch instead of granting the whole workspace — least privilege, enforced outside the prompt. Second, the study decouples spend from quality: API cost spans over 100 times while the overall score spans under 4 times, with the cheapest open models on the Pareto frontier. The manager need not be the expensive model; what it must do well is grant and route, not reason hardest.
Source
- ClawArena-Team: Benchmarking Subagent Orchestration and Dynamic Workflows in Language-Model Agents — https://arxiv.org/abs/2606.31174