frontisai/awesome-self-improving-agents

We welcome issues and pull requests for missing work on harness agents, agent skills, memory, self-improvement, agent RL, evaluation, and safety.

News

• [2026-06-25] 🎉 Release: Our survey is now available on OpenReview.

Citation

If you find this survey or paper list helpful, please cite our work:

@article{jiang2026selfimprovingagents,
  title={Self-Improving Agents in the Era of Experience: A Survey of Self- to Meta-Evolution},
  author={Che Jiang and Jincheng Zhong and Yu Fu and Kai Tian and Junlin Yang and Kaikai Zhao and Yuchong Wang and Tianwei Luo and Weizhi Wang and Yuxin Zuo and Guoli Jia and Xingtai Lv and Dianqiao Lei and Sihang Zeng and Yuru Wang and Zhenzhao Yuan and Xinwei Long and Ermo Hua and Can Ren and Xin Jiang and Shulei Xie and Yuanchun Zheng and Youbang Sun and Biqing Qi and Ning Ding and Kaiyan Zhang and Bowen Zhou},
  journal={OpenReview Archive},
  year={2026},
  url={https://openreview.net/pdf?id=IUltZSgLMm}
}

Contents

• Overview
• Paper List
  • Foundations and Surveys
  • Harness and Runtime Architecture
  • Skills and Skill Libraries
  • Memory and Context Management
  • Environments, Tools, and Runtime Feedback
  • Agent RL and Continual Learning
  • Meta-Agents and Evolution Orchestration
  • Evaluation and Benchmarks
  • Safety and Governance
• Acknowledgment
• Star History

Overview

This repository collects papers, systems, benchmarks, and resources for studying how deployed agentic AI systems become more capable after deployment.

We organize the landscape around the harness agent: a deployed runtime system whose behavior is jointly shaped by a base model, a mutable harness, a user-facing interface, and an environment-facing interface. Under this view, self-improvement first appears as fast runtime adaptation over external surfaces such as skills, memory, context, tools, and execution environments. Repeated experience may later be consolidated into model parameters through reinforcement learning, fine-tuning, or continual learning.

We organize the survey into four parts:

1. Paradigm shift: from task-bounded tool-use loops to deployed harness-centered runtime systems.
2. External path: skills, memory, context, tools, and environments as editable runtime adaptation surfaces.
3. Parameter path and meta-evolution: agent RL, continual learning, and meta-agents for durable learning and update orchestration.
4. Conditions and limits: evaluation, safety, governance, and open problems for reliable post-deployment improvement.

Paper List

This paper list follows the references cited by the LaTeX manuscript. It currently includes 331 unique cited entries from 379 unique manuscript citation keys and 379 cited BibTeX records.

Every row includes a date, display name, title, and at least one public source badge. Cited entries whose public source URL still needs verification are omitted from this table until complete metadata is available (44 currently omitted; 4 duplicate cited records collapsed).

Foundations and Surveys

Harness and Runtime Architecture

Skills and Skill Libraries

Memory and Context Management

Environments, Tools, and Runtime Feedback

Agent RL and Continual Learning

Meta-Agents and Evolution Orchestration

Evaluation and Benchmarks

Safety and Governance

Acknowledgment

This repository is maintained by the FrontisAI and Tsinghua University survey team. Its README structure follows the public awesome-list style of TsinghuaC3I/Awesome-RL-for-LRMs.

Star History