A community and research initiative founded by researchers from Oxford, Columbia, McGill, and more.
We study how capable agents coordinate across information, ownership and control boundaries, and what it takes for that coordination to compound rather than dissolve as networks grow.
Research
Papers from the initiative and from members of its community.
Fig. The RAC cycle. Each level, N0 to R3, adds one coordination capability to the last.
We compare native workflow control with runtime agent coordination (RAC) in multi-agent AI scientists, holding agents, models, tools, permissions and budget fixed so that only the coordination mechanism changes.
Communication Failure in Distributed Runtime Agent Coordination
Fig. Where communication breaks: disclosure, delivery, reading, action. A failure is classed by the first transition that fails.
How does communication fail when agent teams coordinate at runtime without a central orchestrator? Across 746 annotated runs of two to thirty-two agents on research and coding tasks, the harness shapes communication failure more than the model does, failures per post fall as teams grow, and most are repaired later within the run.