Upgrading every agent in a large language model–based multi-agent system can be prohibitively expensive, both in compute and in engineering effort. A new paper on arXiv proposes a way to decide which single agent deserves attention first, using influence functions to estimate each agent's contribution to overall system performance.

The method, called IF-MAP, is inspired by influence-function techniques from robust statistics. Rather than treating all agents equally, it ranks them by how much a small change to one agent would affect the system's final output. This lets developers allocate limited improvement resources where they will have the greatest system-level effect.

The paper's abstract does not report experimental results or compare IF-MAP with other prioritization strategies. It only introduces the problem and the proposed approach, so the practical gains remain to be demonstrated. Still, the framing addresses a real bottleneck in multi-agent development: knowing which component to improve when you cannot improve them all.