Self-Evolving Agent Harness
Chuyang Wei, Haoxiang Guan, Minghang Zhu, Kexin Zhou, Jiyan He, Xiren Zhou, Huanhuan Chen, Yifei Shen
Zhongguancun Academy · Zhongguancun Institute of Artificial Intelligence · University of Science and Technology of China
Published October 13, 2026 · on the feed October 16, 2026 · ▲ 195 · arXiv 2610.18813
The abstract
Harness self-evolution revises an agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs. The framework separates update targets from proposal and admission mechanisms. It distinguishes changes to interpreted state from revisions to tools and code, repository-level control logic, and plugin composition. We characterize search through evolutionary optimization and examine its reliability through continual PAC-learning. Under explicit assumptions, the analysis yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh, independent validation data. We examine how persistent harness revisions affect knowledge sharing and organizational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate model--harness co-design, distributed execution, and plugin reuse. Reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges, and consequential actions. We provide a comprehensive 66-page discussion of related work, theory, algorithms, evaluation, applications, and the broader impact of Harness RSI, together with continuously updated code and supplementary materials, at https://github.com/zgcagi/Harness-RSI.
The Brief
2:25A short podcast summary of the paper. Written and voiced by AI.
Transcript
Self-Evolving Agent Harness. By Chuyang Wei and seven others, from Zhongguancun Academy, the Zhongguancun Institute of Artificial Intelligence, and one other institution. Published on the thirteenth of October. Its code repository is public but slight: five source files.
An agent is more than its model. Around the weights sits a harness: the prompts, the memory, the workflows, and the executable components that make up its software scaffolding. Harness self-evolution is the practice of revising that scaffolding while the model weights stay fixed. The authors treat it as a feedback-driven process, and they formalize it as one in which accepted revisions persist from run to run, so that the agent carries its changes forward.
The framework is organised around a separation. What gets updated is kept apart from how a revision is proposed and how it is admitted. Within the targets of an update, it draws distinctions between changes to interpreted state, revisions to tools and code, changes to the control logic at the level of the repository, and the composition of plugins. Search over revisions is characterised through evolutionary optimization. Its reliability is examined through continual PAC-learning, and under explicit assumptions that analysis yields three results: an expected generalization bound, a retention bound that is conditional on budgets for past regressions, and a rule for admitting a revision with high probability, using fresh, independent validation data.
Beyond the formal side, the authors look at what persistent revisions do to the settings around the agent: to knowledge sharing and organisational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate three patterns: designing the model and harness together, distributed execution, and reusing plugins.
One thread runs through all of it. Reliable deployment, they argue, requires validation that never stops, because task distributions and evaluators keep changing. It also requires independent safeguards for learning, for runtime privileges, and for consequential actions.
The paper is long by design, a sixty-six page treatment of related work, theory, algorithms, evaluation, applications and broader impact, which the authors call Harness RSI. Code and supplementary materials are provided and continuously updated. The thing to hold onto is that a harness which rewrites itself needs its own gatekeeping: each change must earn admission against fresh evidence, and keep earning it.
Paper Quest · 7 chapters · about 31 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Agents keep meeting new tasks after launch. Retraining is not always an option.
You run an agent in the field. Its tasks and environment keep changing after deployment, and it needs a way to keep adapting.
The road ahead
- 1 · The problem ~6 min A deployed agent faces tasks that keep changing, and retraining its model is often costly or simply not possible.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Keep the model's weights fixed and let feedback revise the software around it, keeping only revisions that pass fresh, independent checks.
- 4 · Choose your own research ~3 min
- 5 · How it works ~3 min
- 6 · What they found ~3 min
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.