CosmoEvolve: A Virtual Research Lab as a Hierarchical Multi-Agent System

Abstract: We present CosmoEvolve, an open, domain-general framework that instantiates a virtual research laboratory, consisting of one principal-investigator (PI) agent and a community of student scientist agents, inside a single Python process. Unlike fixed research pipelines, CosmoEvolve leaves the ordering of scientific actions emergent: at every round the PI observes a summarised lab state and selects an action from a finite discrete action space with six elements (group meeting, individual meeting, task assignment, paper request, symposium, and wrap-up), while students execute the selected action through a tool-calling LLM loop that delegates concrete work to read-only and write-enabled subagents. We describe each CosmoEvolve agent as a tuple of LLM backbone, context, tool set, policy, action space, and internal state; describe the two main operating modes (the bounded lab session and the asynchronous lab-continuous mode with a PI thread and parallel student threads synchronised by per-student locks); and present the collaboration primitives (a shared discussion thread acting as a blackboard, sequential group-meeting rollouts, parallel peer review, a shared artifact store, and cross-session memory and skill evolution) that turn the collection of agents into a laboratory that improves itself across runs. We further document the system's tool management, context engineering, and agent visibility architecture in sufficient detail to be reproduced.