The agent field in skill definitions routes task execution to named agent types, indirectly selecting model capability and inference cost.

How it works

When a skill includes agent: {type} in its front matter, the Claude Code runtime passes this value to the Task tool, which resolves it to a specific agent implementation. The field accepts Task tool agent type names—Explore, Plan, general-purpose, Bash, claude-code-guide, and others—not model names directly. Invalid values (e.g. haiku, sonnet) are silently ignored, defaulting to the parent model and capability tier of the invoking session.

What the test found

A test on 2026-01-07 created .claude/skills/test-agent-field/SKILL.md with varied agent field values and observed the resulting model assignments:

The test confirmed that agent type names map to inferred model capability, not the other way around. Explore is documented as a "fast agent" and consistently resolved to Haiku; Plan is documented as a "software architect" and resolved to Opus.

Why it matters

The agent field enables cost-optimized task routing within a single skill suite. A research or exploration skill can declare agent: Explore to run on cheaper inference; architecture or synthesis work can request agent: Plan for deeper reasoning. When combined with context: fork, this allows isolated, single-use agent instances to run at their natural cost tier, rather than forcing all tasks through the parent session's model.

Caveats

Silent failure on invalid agent types means typos pass undetected. The field is only useful if the runtime knows the agent type name; undocumented or deprecated agent types will default silently rather than error. No validation occurs at skill definition time.