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

Harness-MU is introduced as a model-agnostic framework designed for multi-user LLM agents, addressing governance challenges inherent in collaborative workflows. It enforces deterministic governance constraints through execution hooks, decoupling safety from language generation, and achieves significant improvements on the Muses-Bench benchmark, with utility scores enhanced by 0.28–0.39 and instruction-following accuracy increased by up to 48.9 percentage points across multiple models. This framework is crucial for practitioners focusing on secure and compliant interactions in multi-principal environments.