Most benchmarks for multi-agent AI systems assume that agents will cooperate and that the environment stays consistent. But real-world deployments often involve competing goals, contradictory instructions, or other forms of friction. A new arXiv preprint argues that this gap matters, and introduces CLASHarness, a benchmark designed to fill it.
CLASHarness focuses on vision-language models (VLMs) used in multi-agent computer-use workflows. The benchmark's key contribution is a generation pipeline that lets researchers create conflict scenarios in a controlled way, while also making it possible to verify what actually happened. That combination is meant to move testing beyond the usual cooperative assumptions.
The paper is a single source and does not compare against other benchmarks in the abstract, so the claims here are limited to what the authors describe. Still, the work points to a growing recognition that AI agents need to be evaluated under realistic conditions, including when their objectives clash.