Google adds Agent Skills to Genkit Go, enabling on-demand expertise loading

Google has extended its Genkit framework to support Agent Skills in Go, addressing a fundamental challenge in large language model-based applications: managing context window size when agents handle increasingly complex workflows.

The problem with loading everything upfront#

As LLM agents take on more responsibilities, stuffing every standard operating procedure, reference guide, and document into a persistent context window becomes unsustainable. This approach consumes excessive tokens, dilutes the model's attention, and increases error rates. Agent Skills solve this by packaging specialized knowledge into discoverable capabilities that activate only when the agent determines they're needed.

How progressive disclosure works#

The Agent Skills standard relies on progressive disclosure—information surfaces to the model only at the moment it's required. Each skill is structured with a SKILL.md file containing two parts: frontmatter (metadata and description) plus body content (actual instructions). Skills can also bundle supporting files like scripts or reference documentation.

When an agent starts a conversation, Genkit loads only the skill's frontmatter into the system prompt—just enough information for the model to decide whether that skill applies. As the task evolves and the agent identifies a need for a particular skill, the full SKILL.md body and any bundled resources load into the active context.

Three key advantages#

Token efficiency: Detailed instructions load only when a skill activates, not upfront.

Lean infrastructure: Skills are simple markdown bundles distributed like code, requiring no additional configuration.

Built-in resource bundling: Skills can include Python, Node.js, or Bash scripts that agents execute, ensuring access to the right tools for the job.

How it integrates with Genkit#

Agent Skills build on Genkit's middleware architecture, which intercepts critical model lifecycle phases through hooks:

• Model Wrapper – fires per API call, handling retry/fallback/caching logic
• – fires per tool execution, supports parallel calls