An Open General Evolutionary Architecture for
Recursive Self-Improvement of LLM Agent Harness

Abstract

Large language model (LLM) agent behavior depends not only on model weights but also on its harness: instructions, tool definitions, memory, workflows, context management, and execution policies governing environment interactions. Yet existing agent optimization workflows lack general mechanisms to track, validate, and compare changes to these harness components and rarely enforce reproducible provenance for harness edits. To address this gap, we present GEAR, an open framework for harness optimization with evolutionary search and reproducible evaluation. It decouples per-task execution from cross-task improvement, binds every evaluated candidate to immutable implementation and evaluation identities, and keeps search decisions inspectable and recoverable. The framework adopts a two-level meta-learning abstraction, where an inner task agent produces trajectories and evaluation evidence, while an outer meta agent aggregates cross-task evidence to propose reusable harness modifications. We evaluate GEAR on AutomationBench (Marketing and Support) and Terminal-Bench 2.1. With GPT 5.6 Luna at medium effort, the evolved DeepSeek Harness raises task pass rate from 27.0% to 40.0% on Marketing, from 17.0% to 25.0% on Support, and from 52.8% to 62.9% on Terminal-Bench. At max effort, the evolved harnesses reach 84.3% on Terminal-Bench 2.1 with Luna and 87.6% with GPT 6 Astra. We make GEAR publicly available to enable empirical studies of harness evolution as a fundamental component of LLM agent systems. Code is available at https://github.com/rsi-gear/gear.