MASS searches role prompts and agent wiring as one design problem
Do you tune what each agent says, or how the agents are wired?
Opening
Every multi-agent system is built from two dials: the prompt that declares what an agent does, and the topology that decides how agents interact across the workflow. Teams turn them by hand — rewrite a role, redraw the workflow, run again, see if the score moved — and stop when it feels good enough. The trouble is that the dials are not independent: a prompt tuned in one topology can fail in the next, so hand search settles early and calls the result a principle. This paper treats prompts and topologies as one design space, and builds a search that walks both.
What the paper does
The authors first take the design space apart with an in-depth analysis aimed at understanding the factors behind building effective multi-agent systems, and conclude that prompts together with topologies play critical roles. On that reading they propose Multi-Agent System Search (MASS), a framework that efficiently exploits the complex design space by interleaving its optimization stages, from local to global, from prompts to topologies. Three stages run in sequence: block-level prompt optimization, workflow topology optimization, and workflow-level prompt optimization, each conditioned on the prompts and topologies optimized in the stages before it. A later stage therefore sees the output of the earlier one, and a prompt ends up tuned against the wiring it will actually run in.
Findings
Prompts and topologies are one problem — Role prompts and interaction wiring are usually tuned in isolation, which is why a gain from one rarely survives a change to the other. The paper’s analysis of the design space puts both on the critical path of any effective system.
prompts together with topologies play critical roles in enabling more effective MAS design
Interleaved search instead of one-shot tuning — MASS does not fix prompts first and topology second. Each stage is conditioned on what the previous stage optimized, so local prompt edits and the global workflow co-evolve in a stated order: local to global, prompts to topologies.
a MAS optimization framework that efficiently exploits the complex MAS design space by interleaving its optimization stages, from local to global, from prompts to topologies
The searched systems beat the alternatives — Against a spectrum of existing multi-agent construction methods, the systems MASS produces come out clearly ahead, and the authors close by extracting design principles from those winners rather than stopping at the score.
MASS-optimized multi-agent systems outperform a spectrum of existing alternatives by a substantial margin
Caveat
The authors state no limitation in the abstract. What that abstract also does not give is the size of the margin, the tasks it was measured on, or the search cost the three stages demand, so the result reads as a demonstrated direction rather than a number a team can plan a budget against.
What this means for 5ac
5ac currently sells team architecture as a judgment call; MASS says it is a searched configuration. Concretely: when a customer asks for an agent team, the card body should pin the topology that orchestrates interactions across agents and the role prompts together, then change one and re-measure instead of redrawing the whole graph. The interleaving is the lesson — a role prompt written before the workflow exists is tuned against nothing. On the Kanban board that means a profile’s prompt and the parent-child edges of its pipeline get revised in the same card, and the search stops only when runs across both stop improving.
Source
- Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies — https://arxiv.org/abs/2502.02533