Self-Evolving Agent Harness
An agent can retain changes to its prompts, memory, workflows and code across runs while its model weights stay fixed.
The paper formalizes a feedback process that separates what gets changed from how revisions are proposed and accepted. Under explicit assumptions, it derives bounds on generalization and retention of earlier capabilities, plus an admission rule using fresh, independent validation data. Cases from prior work illustrate model–harness co-design, distributed execution and plugin reuse. The authors say reliable deployment requires continual validation as tasks and evaluators change, alongside independent safeguards for learning, runtime privileges and consequential actions.
HF Daily Papers' note
The paper formalizes a feedback process that separates what gets changed from how revisions are proposed and accepted. Under explicit assumptions, it derives bounds on generalization and retention of earlier capabilities, plus an admission rule using fresh, independent validation data. Cases from prior work illustrate model–harness co-design, distributed execution and plugin reuse. The authors say reliable deployment requires continual validation as tasks and evaluators change, alongside independent safeguards for learning, runtime privileges and consequential actions.
HF Daily Papers' note
score 4