Daily Paper

Harness-MU: A Safe, Governed, and Effective Harness for Multi-User LLM Agents

Harness-MU provides a multi-user governance framework for LLM agent deployments, enabling multiple users to share an agent while maintaining safety boundaries, access controls, and audit trails across concurrent sessions.

Wangxuan Fan, Xiaoyu Nie, Zhongxiang Dai

agentic-aisafety-governancemulti-useraccess-controlaudit

Focus: LLM agents are typically designed for single-user interaction, but production deployments often require multiple users to share an agent — raising questions about session isolation, privilege escalation, and cross-user information leakage. Harness-MU addresses these challenges with a governance layer that enforces user-specific safety boundaries without requiring model retraining.

Key Insights

  • Session isolation under shared context: Multi-user agents risk leaking information from one user’s session into another’s context window; Harness-MU uses structured context partitioning to maintain strict session isolation even when the underlying LLM has a shared conversation history.
  • Dynamic privilege enforcement: User permissions are enforced at the tool-call level, allowing different users to have different access rights to the agent’s capabilities (e.g., read-only vs. write access to a shared database) without modifying the agent’s prompting.
  • Audit trail completeness: Every action taken by the agent is attributed to the requesting user and logged, providing an audit trail necessary for governance and incident response in regulated deployments.

Failure-First Relevance

Multi-user agent governance is directly relevant to the Failure-First multi-agent coordination scenario class. Cross-session information leakage and privilege escalation are specific attack vectors that the Failure-First scenarios should include. The audit trail completeness requirement maps onto the Failure-First data preservation principle — complete logs are necessary not only for safety governance but for reproducing and analysing failure modes.