Living-Harness Is an Interactive-Agent Evolver

2026-07-29

AI Summary

Existing static interactive agents cannot continuously learn from failures, leading to repeated errors. This work proposes Living-Harness, a self-evolving agent framework that dynamically updates procedural knowledge by converting interaction trajectories and evaluation signals into structured evidence to guide subsequent behavior. The framework introduces Evolution-SOP, a standard operating procedure that integrates episodic memory and state-graph dual knowledge representations, enabling continual evolution without modifying tools or base context. Built upon large language models, the system incorporates trajectory abstraction, evidence extraction, memory storage, and state-graph retrieval mechanisms. Evaluated across eight environments in τ²-Bench and MultiWOZ-2.4, it achieves average Pass@1 improvements of 10.07 and 9.91 percentage points, respectively, and the evolved framework states demonstrate cross-model backbone reusability.

Abstract

Large language model (LLM) agents may recover from a failure within an episode or after a retry, yet the same execution failure can recur in later tasks because post-episode feedback rarely revises the persistent harness that guides future interactions. Static harnesses improve reliability through fixed tools, context, memory, and workflow structures, but remain unchanged after deployment. We propose $\textbf{Living-Harness}$, a self-evolving agent harness that converts each completed trajectory and its evaluator signals into posterior evidence for bounded harness updates. Guided by a domain-level $\textbf{Evolution-SOP}$ ($\textbf{S}$tandard $\textbf{O}$perating $\textbf{P}$rocedure), Living-Harness extracts an episode abstraction and structured update evidence, and writes two complementary forms of procedural knowledge: episodic memory that records trigger conditions, failure patterns, and recovery actions, and a state graph that records state nodes, repair edges, and transition rules. The updated harness state is retrieved to guide future interactions, while tools and base context remain frozen, allowing procedural repairs to accumulate across evolution cycles. On eight interactive environments derived from $τ^2$-Bench and MultiWOZ-2.4, Living-Harness improves average Pass@1 over the strongest interactive baseline by 10.07 and 9.91 percentage points, respectively, and supports retrieval-only reuse of the evolved harness state across model backbones.

Problem

• LLM agents
• execution failure
• persistent harness
• procedural knowledge
• interactive environments

Innovation

• Living-Harness
• self-evolving agent
• Evolution-SOP
• procedural knowledge
• harness update