Mycelium: Active shared context graphs turn teams into networked intelligence

Most AI-for-science systems focus on improving a single reasoning process through better models or larger context windows. But real scientific breakthroughs emerge from teams with diverse priors, experimental backgrounds, and tacit knowledge. The challenge is not just scaling models but cultivating networked intelligence—scaling connections between humans and AI so that insights flow to the right person, agent, or instrument. To address this, Sutanay Choudhury and 17 co-authors from Pacific Northwest National Laboratory (PNNL) and other institutions introduce Mycelium, an active shared workspace that functions as a multi-user co-scientist. As human users and AI agents work, the system captures observations and hypotheses, tracks their relationships to the team's evolving model, and routes them automatically to the human or agent who can act on them.

Mycelium was tested in its first empirical evaluation: a biological multi-omics campaign. The system demonstrated that routed shared context could transform a local analytical finding into a cross-expert mechanistic constraint, ultimately shaping an experimental design. The paper also provides a formal computational account of networked intelligence as sparse conditional computation over distributed scientific contexts. This theory distinguishes when a scaled standalone agent can match the network's performance versus when independent expertise and non-mergeable contexts make the network irreducible. By enabling dynamic, context-aware collaboration between multiple humans and multiple AI agents, Mycelium represents a paradigm shift from single-reasoner AI assistants to true team-wide co-intelligence.