Researchers Propose Self-Evolving Agent Harness Framework for Fixed-Model AI Agents

Researchers have introduced a framework for Self-Evolving Agent Harnesses, a method that revises an agent's software scaffolding—including prompts, memory, workflows, and executable components—while keeping the underlying model weights fixed. This approach separates update targets from proposal and admission mechanisms, distinguishing between changes to interpreted state and revisions to tools, code, repository-level control logic, and plugin composition.

The study characterizes the search process through evolutionary optimization and examines its reliability using continual PAC-learning. Under explicit assumptions, the analysis provides:

• An expected generalization bound.
• A retention bound conditional on historical regression budgets.
• A high-probability admission rule using fresh, independent validation data.

The authors discuss the implications of persistent harness revisions for knowledge sharing, organizational workflows, scientific reliability, physical safety, and institutional accountability. The paper includes a 66-page discussion on related work, theory, algorithms, evaluation, and broader impacts, alongside continuously updated code and supplementary materials.