Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

A recent study published on arXiv (2608.08466) presents Hierarchical Self-Improvement (HSI), a novel framework that allows LLM agents to adapt their executable scaffolds, or harnesses, for particular tasks. In contrast to conventional methods that view harnesses as static post-deployment, HSI utilizes distinct harnesses for each task family, which can be swapped through a task-injection seam and modified based on feedback from the environment. The framework functions at three hierarchical levels: a task harness that performs tasks, an evolver that modifies the harness, and a meta-evolver that updates the evolver's strategy code while maintaining a stable outer anchor. A thinking-on/off mechanism separates the evolution of the harness by turning off reasoning during task execution and enabling it during self-modification. This paper is classified as a new announcement and can be accessed via the provided URL.