Skills and Agent Evolution Research
Key Takeaways
- Agents evolving from tool-calling to self-evolution (HyperAgents, Memento-Skills)
- Skills crystallizing as the basic unit of reusable capability
- Enterprise Skills ecosystems forming (MiniMax Office Skills, Claude Code skill sets)
- Self-optimization limitation: optimizes for error rates, not user-specific edge cases
- Key frameworks: HyperAgents (meta-level), Memento-Skills (Read-Execute-Reflect-Write), Deep Agents
- 智能体正从工具调用向自我演化演进（HyperAgents、Memento-Skills）
- 技能正在固化为可复用能力的基本单元
- 企业技能生态正在形成（MiniMax Office Skills、Claude Code skill sets）
- 自我优化的局限：针对错误率进行优化，而非针对用户特定的边缘情况
- 关键框架：HyperAgents（元级）、Memento-Skills（读取-执行-反思-写入）、Deep Agents
Summary
This research compilation (March 26, 2026) explores how Skills are becoming the core mechanism for agent capability growth. Four major trends: agents evolving to self-evolution, Skills as the basic unit of reusable capability, enterprise Skills ecosystems forming, and fundamental limitations of agent self-optimization discovered.
The tension between self-evolution promise and practice: while Memento-Skills elegantly turns failures into learning signals, critics note that agents optimize for aggregate error rates rather than edge cases users care about.
自我进化承诺与实践之间的张力：尽管 Memento-Skills 优雅地将失败转化为学习信号，但批评者指出，智能体优化的是总体错误率，而非用户关注的边缘案例。