这篇 working survey 试图回答一个问题：AI Agent 的“自我改进”什么时候只是反复尝试，什么时候已经形成可验证、可保留、可回滚的系统过程？当前版本连接 196 篇论文展示集、686 个 GitHub 原始捕获、300 份项目/模型卡分析和 97 个公开痛点信号；这些数字说明证据来源，不等同于完成度或排名。
Working survey draft on AI self-evolution: taxonomy, methods, frameworks, evaluation, and future directions. 8 chapters + appendix. Treat sections on evaluation, pain points, and framework scoring as provisional until their evidence audit is complete.
当前发布的是 44 页中文伴读版，便于快速阅读核心框架、方法分类、评估体系和未来方向。长版中文稿仍是校对中的同步版本，不作为公开最终稿引用。
这里不再只放旧的站点装饰图；每张图都对应一个 paper chapter 和一个公开 evidence boundary。
[INFERRED] 2026-05-26 自动关键词初分类（196-paper display snapshot）。它用于说明当前语料倾向，后续需要由 paper review、纳入/排除标准和人工复核校正。
| Framework | Adoption | Observability | Production Fit | Self-Evolution Fit | Safety/Governance | Status |
|---|---|---|---|---|---|---|
| LangChain/LangGraph | ● 5/5 | ○ 3/5 | ● 5/5 | ○ 3/5 | ● 3/5 | [INFERRED] |
| CrewAI | ● 5/5 | ● 4/5 | ● 4/5 | ○ 2/5 | ○ 2/5 | [INFERRED] |
| AutoGen | ● 5/5 | ● 4/5 | ● 4/5 | ○ 2/5 | ○ 2/5 | [INFERRED] |
| OpenAI Agents SDK | ● 5/5 | ○ 1/5 | ● 5/5 | ○ 2/5 | ● 3/5 | [INFERRED] |
| EvoAgentX | ● 5/5 | ● 4/5 | ● 4/5 | ● 5/5 | ○ 2/5 | [INFERRED] |
| DGM/OpenEvolve | ● 5/5 | ● 4/5 | ● 5/5 | ● 5/5 | ○ 2/5 | [INFERRED] |
[INFERRED] 这些 1-5 分仍是启发式草稿，基于 repo 信号和痛点分析；还缺少完整 rubric、打分人、反例、reviewer disagreement 和逐项来源，因此不能当学术排名引用。
自进化研究前沿（ADAS、DGM、Absolute Zero、Reflexion 等）不一定会同步反映在 GitHub Star 数中。这里的读者提示很简单：Star 只能作为 adoption prior，不能替代论文证据、可运行实现、评估协议和失败记录。
| Research Method | Paper Signal | GitHub Signal | Reader Interpretation |
|---|---|---|---|
| Architecture Search (ADAS) | ICML 2024 | repository dependent | Check method evidence before adoption signal. |
| Godel Machine (DGM) | arXiv 2025 | repository dependent | Inspect retained changes, evaluator, and rollback evidence. |
| Self-Play Reward (Absolute Zero) | arXiv 2025 | repository dependent | Do not infer maturity from stars alone. |
| Reflection (Reflexion) | NeurIPS 2023 | paper/code ecosystem | Separate research influence from production readiness. |
| Dataset | Count / Snapshot | Basis | Limit |
|---|---|---|---|
| Survey-display papers | 196 | 196-paper display snapshot used for the current figure atlas. | Automatic keyword grouping; not a final systematic-review denominator. |
| Paper review files | 2 | Public paper report files exported under site/public/reports/papers. | Coverage is file-count based; individual claims still need source-level audit. |
| GitHub raw captures | 686 | 300 project/model-card analyses and 491 public project reports are currently surfaced. | Raw capture count is not project quality, maturity, or research value. |
| Public pain-point signals | 97 | Semantic analysis of Reddit, Hacker News, and X/Twitter public discussions. | These are public discussion signals, not formal interviews or survey respondents. |
| Site corpus snapshot | 2026-06-05 | Counts generated by the public corpus/index pipeline. | Use this as a dated snapshot; hourly metadata may update downstream pages. |
Mom Test 97 痛点来源：Reddit 47、Hacker News 36、X/Twitter 14。它们是公开讨论信号，不是受控访谈样本。
Definition, scope, contributions, boundaries
Five evolution loops and formal design dimensions
Inference-time and training-time self-improvement
Code evolution, algorithm discovery, agent self-evolution
Benchmarks, process metrics, disclosure standard
Agent frameworks and infrastructure gaps
User pain points and production constraints
Open problems, roadmap, conclusion
Projects, papers, benchmarks, star visibility signals
可作为 working draft 引用：定义框架、五类进化回路、主要证据边界和当前公开 PDF；引用时必须标注版本日期与 provisional status。
谨慎引用：方法族比例、框架 1-5 分、项目覆盖数字和痛点统计，它们仍是 dated snapshot。
不要引用为最终结论：Evolve-AGI worksheet、Star/Rank 页面、framework radar 和 corpus count。
引用数字时必须带日期：当前公开语料快照为 2026-06-05，部分站点元数据会继续更新。
需要继续补强：逐项 citation mapping、benchmark source audit、中文长稿同步和英文网页镜像的图表逐项说明。