Google has introduced Agent Skills for Genkit Go, a pattern for packaging specialized agent instructions into SKILL.md bundles that are loaded only when a task calls for them.

The update is aimed at a familiar problem in agent systems: prompts grow as teams add more procedures, policies, examples, and tool-specific guidance. Google describes Agent Skills as a progressive disclosure mechanism. At startup, Genkit exposes only each skill's frontmatter metadata to the agent. If the conversation matches a skill, the middleware can then load the full body of the relevant instructions.

How it works

The Google Developers Blog says the flow has three stages: discovery, activation, and execution. During discovery, Genkit scans configured SkillPaths for SKILL.md files and adds their metadata to the system prompt. During activation, a matching task triggers the full skill body. During execution, the agent can use those more detailed instructions to complete the task.

Google's example shows the feature configured through Genkit Go middleware with ai.WithUse(&middleware.Skills{SkillPaths: []string{"./skills"}}). The linked demo repository includes a simple recipe example, but the post frames the pattern for larger production agents that need modular instructions without keeping every possible workflow in context at once.

Why it matters

Agent frameworks are converging on more structured ways to manage context, tools, and domain knowledge. Agent Skills gives Go developers a repo-native packaging model for reusable instructions, while keeping the always-on prompt smaller.

The release does not make the agent decide reliably by itself; teams still need to design the metadata and skill contents carefully. But it gives Genkit Go users a first-party path for separating baseline behavior from task-specific expertise.