Definition
Plain language
A multi-agent AI research system where parallel agents share notes through a trained communication layer.
As stated in the literature
A long-horizon agent framework from Renmin and BAAI that connects parallel ReAct agents through a trained hub model that compresses scratchpads into structured notes and serves targeted readouts on demand, optimizing the communication layer with RL while freezing the agents.
Why it matters: Naive multi-agent setups drown in each other's transcripts, and a trained communication layer is one path to making real collaboration scale.
For example, three search agents each exploring a research question can share their findings through AgentFugue's hub instead of dumping raw scratchpads into each other's contexts.
Heard on the show
“The paper is called "AgentFugue: Agent Scaling for Long-Horizon Tasks through Collective Reasoning," out of Renmin University and the Beijing Academy of Artificial Intelligence.”Episode 083 — Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves