Mozi: Governed Autonomy for Drug Discovery LLM Agents

2026-02-27 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Drug Discovery · Depth: Advanced, extended

Summary

Mozi is a dual-layer architecture designed to overcome unconstrained tool-use governance and poor long-horizon reliability in large language model (LLM) agents for drug discovery. It integrates generative AI flexibility with computational biology rigor. Layer A, the Control Plane, establishes a governed supervisor-worker hierarchy, enforcing role-based tool isolation, limiting execution to constrained action spaces, and driving reflection-based replanning. Layer B, the Workflow Plane, operationalizes canonical drug discovery stages, from Target Identification to Lead Optimization, as stateful, composable skill graphs. This layer incorporates strict data contracts and strategic human-in-the-loop (HITL) checkpoints to ensure scientific validity at high-uncertainty decision boundaries. Mozi provides built-in robustness and trace-level audibility, demonstrating superior orchestration accuracy on PharmaBench and generating competitive in silico drug candidates in end-to-end therapeutic case studies for Crohn's disease, Parkinson's disease, and sepsis.

Key takeaway

For AI Engineers and Research Scientists developing LLM agents for high-stakes scientific domains, Mozi's dual-layer architecture offers a blueprint for enhancing reliability and auditability. You should consider implementing hierarchical agent systems with hard-coded tool filtering and workflow-native skill graphs to manage long-horizon tasks, ensuring reproducibility and mitigating error accumulation. Integrate human-in-the-loop checkpoints at critical decision boundaries to maintain scientific validity and enable expert intervention, transforming agents into trustworthy co-scientists.

Key insights

Mozi unifies LLM flexibility with scientific rigor via dual-layer governed autonomy for drug discovery.

Principles

• Free-form reasoning for safe tasks, structured execution for long-horizon pipelines.
• Hierarchical supervision enforces role-based tool isolation and constrained action spaces.
• Workflow graphs with state contracts mitigate hallucination and error propagation.

Method

Mozi employs a Control Plane (Layer A) for hierarchical supervision and reflection-based replanning, and a Workflow Plane (Layer B) that uses stateful skill graphs with HITL checkpoints for canonical drug discovery stages.

In practice

• Implement role-based tool access control for LLM agents.
• Use stateful skill graphs for complex, multi-stage workflows.
• Integrate human-in-the-loop checkpoints at high-uncertainty decision points.

Topics

• LLM Agents
• Drug Discovery
• Governed Autonomy
• Multi-Agent Systems
• Computational Biology

Best for: AI Engineer, AI Scientist, Research Scientist

Related on AIssential

• SMDD-Bench: Can LLMs Solve Real-World Small Molecule Drug Design Tasks? — SMDD-Bench reveals current LLMs solve only 40.2% of complex drug design tasks, underscoring significant challenges in autonomous scientific discovery.
• Stanford researchers will discuss their agentic 'scientists' that are on course to reshape drug discovery at VB Transform 2026 — Agentic AI systems with hierarchical orchestration and continuous context can significantly improve the efficiency of complex, multi-stage processes like drug discovery.
• Dual-Agent Framework for Cross-Model Verified Translation of Natural-Language Protocols into Robotic Laboratory Platform — A dual-agent framework translates natural-language lab protocols into robotic commands with LLM-verified self-correction.
• Agent libOS: A Library-OS-Inspired Runtime for Long-Running, Capability-Controlled LLM Agents — Agent libOS provides a secure, auditable runtime for long-running LLM agents by making runtime primitives the authority boundary.
• From Data to Theory: Autonomous Large Language Model Agents for Materials Science — Autonomous LLM agents can perform end-to-end scientific theory development, from equation selection to code execution and validation.
• Accelerating pharmaceutical research with a multi-agent AI system — Multi-agent AI systems can transform fragmented data into actionable insights, significantly accelerating complex research processes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.