Mozi: Governed Autonomy for Drug Discovery LLM Agents

A research team led by He Cao has introduced Mozi, a novel architecture designed to bring governed autonomy to large language model (LLM) agents operating in the high-stakes domain of drug discovery. The system directly addresses critical bottlenecks that have prevented the reliable deployment of AI agents in pharmaceutical pipelines: unconstrained tool-use governance and poor long-horizon reliability. Mozi's core innovation is a dual-layer design that bridges the flexibility of generative AI with the deterministic rigor required in computational biology, aiming to prevent the multiplicative compounding of early-stage errors that typically derail autonomous scientific workflows.

Mozi's architecture consists of two key planes. The Control Plane establishes a governed supervisor-worker hierarchy that enforces role-based tool isolation, limits execution to constrained action spaces, and drives reflection-based replanning. The Workflow Plane operationalizes canonical drug discovery stages—from Target Identification to Lead Optimization—as stateful, composable skill graphs. This layer integrates strict data contracts and strategic human-in-the-loop (HITL) checkpoints to safeguard scientific validity. Evaluated on the PharmaBench benchmark, Mozi demonstrated superior orchestration accuracy over existing baselines. In end-to-end case studies, it successfully navigated massive chemical spaces, enforced stringent toxicity filters, and generated highly competitive in silico drug candidates, effectively evolving the LLM from a fragile conversationalist into a traceable and reliable co-scientist.