Large language model–based multi-agent systems are increasingly used to solve complex tasks by splitting work among specialized agents and combining complementary expertise. But according to a new arXiv preprint, these systems and their evaluations largely rely on an implicit assumption: that every agent performs flawlessly.

The paper, titled "Beyond Perfect Teammates: Multi-Agent Collaboration with Unreliable Colleagues," argues that this assumption is unrealistic. In practice, agents may give wrong answers, miss subtasks, or fail to communicate—yet most existing frameworks do not account for such failures.

Because the abstract is truncated, the full method and results are not available from the source. What is clear is the motivation: making multi-agent collaboration robust to unreliable colleagues is an open problem, and this work appears to target that gap directly.