MemoHarness: Adaptive AI Agent Control Learns from Experience

The "agent harness" is the external control layer that transforms a base Large Language Model (LLM) into an executable agent, managing critical functions like context, tools, orchestration, and memory. While harness design profoundly impacts agent behavior, most optimization efforts focus on narrower aspects like prompts. This paper introduces MemoHarness, a novel framework for adaptive harness optimization that learns directly from its own operational experience. MemoHarness breaks down the agent harness into six distinct, editable control dimensions. It maintains a dual-layer experience bank, storing both case-specific diagnoses and distilled global patterns. This accumulated experience allows MemoHarness to adapt the harness configuration to each new test case, retrieving relevant insights without requiring real-time labels, feedback, or additional search. Evaluations across diverse benchmarks, including shell-agent tasks, code generation, and analytical reasoning, demonstrate that MemoHarness consistently outperforms fixed harness configurations. The framework also shows selective transferability to unseen test suites and different base models. Furthermore, its ability to cache retrieved experience can maintain cost-competitiveness. These findings suggest that leveraging execution experience is a practical way to build more adaptive and efficient agent harnesses.