Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.19128.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

CLASHarness is a controllable benchmark of 600 multi-round tasks for evaluating how vision-language agents in computer-use settings handle conflict from peer agents (outdated states, contradictory instructions, shared lock contention). Evaluation of 15 frontier models shows three bottlenecks: weak models follow misleading peers, fixing conflicts after detection is harder than spotting them, and 58%+ of closing actions fail due to coordination gaps.

Quantifies coordination as the universal bottleneck in multi-agent computer use, with concrete failure rates across 15 models. Useful for anyone designing or selecting agent systems that share state.

Cited by

No citations on record.