Self-Evolving Agent Harness
The paper defines “harness self-evolution” as changing an agent’s scaffolding while leaving the model weights fixed.
Accepted revisions can persist across runs, covering prompts, memory, workflows, executable components, repository logic, and plugin composition. The authors separate what gets updated from how changes are proposed and admitted. They frame reliability with evolutionary search and continual PAC-learning, including bounds and an admission rule using fresh validation data. The paper argues deployment needs continual validation and independent safeguards around learning, runtime privileges, and consequential actions. ArXiv · AI/CL/LG's note
Accepted revisions can persist across runs, covering prompts, memory, workflows, executable components, repository logic, and plugin composition. The authors separate what gets updated from how changes are proposed and admitted. They frame reliability with evolutionary search and continual PAC-learning, including bounds and an admission rule using fresh validation data. The paper argues deployment needs continual validation and independent safeguards around learning, runtime privileges, and consequential actions. ArXiv · AI/CL/LG's note
score 5