Self-Evolving Agent Harness: revising scaffolding with PAC-bound guarantees, frozen weights
zgcagi · hf · 2026-10-16
Researchers from Zhongguancun Academy and USTC propose a framework for self-evolving agent harnesses: revising an agent's software scaffolding (prompts, memory, workflows, executable components) while keeping model weights fixed.
- Formalizes a feedback-driven revision process in which accepted revisions persist across runs, distinguishing update targets from proposal/admission mechanisms, state interpretations from tool/code revisions, and repo-level control logic from plugin composition.
- Characterizes search via evolutionary optimization and analyzes reliability through continual PAC-learning, yielding an expected generalization bound, a retention bound conditional on historical budgets, and a high-probability admission rule using fresh independent validation data.
- Discusses implications for knowledge sharing, scientific reliability, physical safety and institutional accountability, with cases of model-harness co-design, distributed execution and plugin reuse.
More from coding & agent
- SERO augments long-horizon agentic RL with executable oracle continuations — Shanghai-AI-Laboratory · 2026-10-16
- Alibaba open-sources MAI-Gym, an all-in-one GUI agent eval and RL training platform — Tongyi-MAI · 2026-10-16
- K8s FM: managing Spotify playlists declaratively with a Kubernetes operator — saheedniyi_02 · 2026-10-16
- What are cloud agent swarms actually for? Real use cases from federal courts to overnight testing — Dan_Jeffries1 · 2026-10-16
- Codex reads across chat windows to self-test whether a skill fired — lxfater · 2026-10-16
- PixelLab tackles ComfyUI pain points: model dedup, image triage, workflow repair — SmartRick137 · 2026-10-16