本文由 AI 分析生成
建立時間： 2026-05-23 來源： https://x.com/realmcore_/status/2039382343581147414
Summary
This technical blog from RandomLabs argues that agent skills are misimplemented across the industry: they are treated as manually-invoked static prompts (slash commands) rather than contextual behaviors that agents activate dynamically based on environment state. The post introduces Slate’s execution model with threads, forking, and orchestration skills as a solution, enabling skill automation and interactive skill use without polluting the main context.
RandomLabs 的技術部落格指出業界普遍誤用 agent skills：大多數系統把 skill 當成靜態 prompt（slash command），而非根據上下文動態啟用的情境行為。文章介紹 Slate 的 thread + forking 執行模型，讓 skill 能作為子執行緒動態執行，支援互動式使用，並透過「orchestration skill」定義 skill 鏈式執行序列。
Prerequisites
- Understanding of LLM agent harnesses and orchestration loops
- Familiarity with context window management
- Basic knowledge of continuation-passing style (for thread model analogy)
Core Idea
Skills should be situational behaviors that compose into larger sequences guided by goal and environment state. Slate implements this via isolated threads (scoped context, permissions, task execution) that can be spawned for each skill invocation. A new forking primitive supports synchronous, interactive skill use (blocking the main orchestrator, allowing user conversation) without the UX problems of async dialogs. Orchestration skills are higher-order skills that reference other skills and define their sequencing.
Results
Skill chaining is released in Slate. The forking model resolves the user-interaction problem for interactive skills (planning, decision-making) that prior approaches (popup dialogs, forced escalation) could not handle cleanly. As of April 2026, forking is in alpha.
Limitations
- Author-stated: forking feature pushed back to alpha as of April 2026 for reliability
- Unstated: the orchestration skill approach still requires humans to author the sequencing logic; autonomous skill composition remains unsolved; performance with many available skills not benchmarked
Reproducibility
- Code: not open-sourced (Slate is a product)
- Datasets: not applicable
- Compute: not applicable
Insights
The analogy between Slate threads and OS processes/continuations (fork, permissions, context-as-memory) is deliberate and illuminating. The “knowledge overhang” concept — the gap between what a model knows how to do and what it actually chooses to do — provides a precise justification for skills: they shift in-distribution behavior toward out-of-distribution knowledge without fine-tuning. The practical observation that models rarely self-activate skills (users typically do it manually) identifies a key gap between the theoretical and actual value of current skill implementations.
Connections
Raw Excerpt
Skills should be situational behaviors that are composed into larger sequences where the goal and environment state guide the usage of said skill. A skill should be something the agent does not something the agent reads.