EvoTeam Moves Multi-Agent Evolution to Team Structure

A new paper on arXiv, EvoTeam: Multi-Agent Collaboration via Evolving Teams, argues that the usual recipe for improving LLM-based multi-agent systems — refining the prompts, skills, and execution strategies of individual agents — leaves collaboration on the table.

The authors state the problem plainly: "improving individual agents does not necessarily lead to better team collaboration." Their proposal, EvoTeam, is described as "a collaboration-centric evolution framework that jointly evolves team organization, interaction rules, and individual agent skills through execution experience." Rather than reusing one shared structure across tasks, it "maintains a library of reusable teams, selecting or creating agent teams and guiding collaboration through team-specific interaction rules."

The design choice carrying the claim is how experience is bounded. "Collaboration experience accumulates within compatible team contexts, while individual skills remain reusable across teams," the paper says — a split meant to avoid reusing interaction rules across "incompatible team contexts," which the authors flag as a weakness of prior collaboration-level evolution work.

Validation, as stated, is thin on the page. "Experiments across four benchmarks demonstrate its effectiveness," the abstract reports. The abstract names none of the four, publishes no per-benchmark numbers, and lists no baselines.