EVOCHAMBER: Multi-Agent AI System Evolves Collaboration at Test Time

Researchers have introduced EVOCHAMBER, a training-free framework for multi-agent test-time evolution that operates at individual, team, and population scales. Unlike prior methods that either isolate agent experiences or broadcast symmetrically, EVOCHAMBER enables asymmetric knowledge routing and emergent specialization. The core mechanism, CODREAM (Collaborative Dreaming), is a post-task protocol triggered by team failure or disagreement, where agents collaboratively reflect, distill insights, and route them asymmetrically. This approach addresses the limitation of single-agent evolution, which only evolves context and memory, by also evolving collaboration structures, knowledge flow, and specialization across the agent population. The work is detailed in arXiv paper 2605.11136.