Genkit GoのAgent Skillsでオンデマンドの専門知識を活用するEnable on-demand expertise with Agent Skills in Genkit Go

Google has introduced Agent Skills in Genkit Go, its Go-based framework for building AI-powered applications, giving developers a structured way to supply large language models with specialized expertise only when it is needed. The addition matters because an agent's context window is a scarce and costly resource, and packing it with every possible instruction tends to inflate token consumption, slow responses, and dilute the model's focus.

At the core of the feature is what Genkit describes as a progressive disclosure architecture. Rather than injecting all guidance into a single prompt, developers package specialized instructions, scripts, and reference material into modular bundles built around a SKILL.md file. Each bundle represents a discrete area of competence, and only the frontmatter metadata is surfaced to the model at first. That metadata acts as a lightweight index, allowing the agent to recognize that a skill exists and understand roughly what it does before deciding whether to load its full contents.

The practical payoff is a reduction in context window bloat and lower token usage. Because the detailed body of a skill, along with any bundled scripts and references, remains out of the prompt until the agent actually calls on it, routine interactions stay lean. When a task genuinely requires deeper domain knowledge, such as a particular data-processing routine or a compliance checklist, the agent can pull that skill in on demand. This on-demand model is what the announcement frames as leveraging expertise without paying to keep it resident in the prompt at all times.

The SKILL.md format itself follows a familiar convention. Frontmatter carries structured metadata such as a name and description, while the body holds the fuller instructions the model consumes once the skill is activated. Grouping scripts and reference files alongside that markdown keeps a capability self-contained and portable, which should make skills easier to version, share, and reuse across projects.

Some background helps explain why this design is appearing now. Genkit is Google's open-source framework for developing generative AI features, originally released for JavaScript and TypeScript and later extended to Go, and it provides abstractions for prompts, tools, flows, and model providers. Agent Skills sits alongside existing mechanisms such as tool calling and retrieval-augmented generation, or RAG, which likewise aim to give models access to information beyond their training data. Where RAG typically retrieves passages from a vector store and tool calling exposes callable functions, Agent Skills packages procedural know-how and instructions as reusable units, addressing a slightly different need.

The broader industry has been converging on comparable patterns. The SKILL.md bundle format and the idea of progressive disclosure echo approaches other vendors have promoted for agent development, and standards efforts such as the Model Context Protocol have similarly sought to give agents modular, discoverable access to external capabilities. Viewed in that light, Genkit Go's Agent Skills appears to be Google's take on a recurring problem: how to scale an agent's knowledge without letting its prompt