Networked Intelligence: Active Shared Context Graphs for Human-AI Team Science

Scientific discovery is increasingly a team-based endeavor, yet current AI tools are primarily designed for individual users or isolated agentic workflows. This paper addresses the challenge of coordinating distributed expertise, asking how to build a system that allows scientific context—such as observations, hypotheses, and provenance—to flow reliably between humans, AI agents, and laboratory instruments to accelerate discovery.

The authors introduce Mycelium, a runtime architecture built around an Active Context Graph (ACG). Unlike passive databases, the ACG is an active, directed graph that stores scientific state as typed entries (e.g., observations, hypotheses, findings). The system implements three core operations:

1. Cross-user state routing: Automatically pushing relevant findings to the researchers or agents who need them.
2. Persistent knowledge modeling: Maintaining a shared, evolving state across sessions.
3. Provenance-bounded propagation: Ensuring every claim is strictly linked to its empirical evidence and derivation history.

Researchers interact with Mycelium through standard AI chat interfaces, which are connected to the backend via the Model Context Protocol (MCP). The system uses dynamic code generation for autonomous analysis and includes automated error recovery to ensure reproducibility.

Mycelium was evaluated in a multi-omics microbial phenotyping campaign involving three researchers with complementary expertise. The system successfully reconciled isolated findings—such as a regulatory anomaly in one thread and a proteomic signal in another—into a unified mechanistic model.

When benchmarked against standalone AI agents, Mycelium demonstrated superior performance in the "evidence-to-action" pipeline. While standalone agents could generate individual findings, Mycelium produced a significantly higher number of actionable, evidence-grounded experimental proposals. The networked approach effectively expanded the breadth of scientific discovery without sacrificing the specificity of the resulting artifacts, proving that coordination architecture is as critical as model scaling for complex team science.