Mass Framework: Boosting Multi‑Agent Design with Smarter Prompts & Topologies

In multi‑agent systems (MAS), designing effective prompts and topologies is challenging because individual agents can be sensitive to prompts and manual topology design requires extensive experiments.

To automate the design process, Google and Cambridge University analyzed the design space, finding that prompt design has a significant impact on downstream performance, while effective topologies occupy only a small portion of the search space.

Experiments with Gemini 1.5 Pro on mathematical problems show that better prompts combined with more compute yield higher accuracy, and that not all topologies positively affect MAS performance.

The authors propose the Mass framework, which optimizes MAS through three stages:

Block‑level (local) prompt optimization: Optimize prompts for agents within each topology block.

Workflow topology optimization: Optimize the workflow topology within a pruned topology space.

Workflow‑level (global) prompt optimization: Perform global prompt optimization on the best‑found topology.

The Mass framework interleaves prompt and topology optimization in a customizable MAS design space, discovering effective designs as illustrated on the right.

Experimental results using Gemini 1.5 Pro and Flash models compare Mass against methods such as Chain‑of‑Thought, Self‑Consistency, Self‑Refine, Multi‑Agent Debate, ADAS, and AFlow, showing:

Performance improvement: Mass outperforms existing methods on multiple tasks, with an average gain of over 10%.

Importance of staged optimization: Each stage contributes to performance gains, confirming the necessity of local‑to‑global optimization.

Co‑optimization of prompts and topologies: Simultaneous optimization yields better results than optimizing either alone.

Cost‑effectiveness: Mass achieves stable, efficient performance improvements with higher sample efficiency and lower cost compared to other automated design approaches.

For further reading, see the original paper: https://arxiv.org/pdf/2502.02533