Adding agents does not automatically improve a multi-agent system: collaboration rules and task instructions can compete inside the same prompts. The useful shift here is to optimize those two jobs separately, so a change to how agents coordinate can be shared across the team without rewriting each agent’s task procedure.
Key Findings
- Across 11 benchmarks in four domains, applying the framework improved four existing prompt optimizers, suggesting it complements their search strategies rather than replacing them.
- With PE2 on document-understanding systems, average gains were 4.3 points for MDocAgent and 2.7 for DocAgent across five benchmarks. With MASPO, results improved on all five math and code benchmarks.
- In a controlled document-understanding study, the approach improved optimization-budget efficiency and produced steadily rising scores as the tested system grew from one to five agents. These scaling results are specific to that setup, not a guarantee for every multi-agent system.
How the Method Works
A shared coordinator prompt defines collaboration behavior—such as sharing findings and handling disagreement—while each executor prompt specifies its agent’s assigned work. The combined prompts are optimized against the system-level performance objective, but role-specific guidance tells the optimizer whether to refine coordination or execution. Existing optimizers still perform the underlying search.
Who It’s For
A good fit for researchers tuning multi-agent systems who want coordination rules to remain consistent while agent tasks vary, especially when adapting an existing prompt optimizer. Look elsewhere if you need a new optimizer, a guarantee of improvement outside the tested settings, or evidence that prompt optimization alone solves coordination failures; the reported gains depend on task, agent setup, and validation.