Multi-agent system failures driven by specification errors

Research into multi-agent systems (MAS) indicates that 79% of failures are caused by poor specification and coordination rather than model capability limits. According to the MAST taxonomy developed by UC Berkeley researchers, failures are primarily categorized into System Design Issues (44.2%), such as step repetition and failure to obey task specifications, and Inter-Agent Misalignment (34.4%), which includes reasoning-action mismatches.

While the industry has focused on runtime enforcement and governance—exemplified by Broadcom AgentMinder and Microsoft MXC—critics suggest these efforts address only the symptoms of the problem. The data suggests a need for better foundational engineering in how agent roles and tasks are defined.

In a related development, the HarnessDev project, a joint effort by the ByteDance Seed team and various universities, is exploring whether large language models (LLMs) can autonomously build and iterate on agent harness systems. These harnesses serve as the runtime control layer for agents, managing execution loops and state persistence. While experiments show LLMs can generate functional harness code, challenges remain regarding logical module activity, portability, and runtime token overhead.