Living-Harness Is an Interactive-Agent Evolver

Large language model (LLM) agents often suffer from recurring execution failures because corrections made during a single task instance are typically discarded once the episode ends. The authors investigate how to transform post-episode feedback into persistent procedural knowledge that can guide an agent's behavior in future, similar tasks without requiring full model retraining.

The authors propose Living-Harness, a framework that separates an agent's fixed execution resources (tools and base context) from its evolving procedural state. The system operates through a rollout-evaluate-update loop:

1. Rollout: The agent interacts with the environment using current procedural knowledge retrieved from its memory and state graph.
2. Evaluate: Upon completion, an evaluator provides feedback on the trajectory.
3. Update: A domain-specific Evolution-SOP (Standard Operating Procedure) processes the trajectory and evaluation signals to extract structured evidence. This evidence is used to update two components: episodic memory (storing trigger conditions and recovery actions) and a state graph (storing transition rules and repair edges).

This design ensures that procedural repairs—such as knowing when to invoke a specific tool or how to handle a specific error—persist across episodes, while the core operational boundaries of the agent remain stable.

Living-Harness demonstrates significant improvements in task reliability across eight interactive environments derived from τ²-Bench and MultiWOZ-2.4. It outperformed the strongest interactive baselines, achieving an average Pass@1 improvement of 10.07 and 9.91 percentage points, respectively. Furthermore, the authors show that the evolved harness state is portable, supporting retrieval-only reuse across different model backbones, which suggests that the procedural knowledge captured is model-agnostic.