Skip to content

Load agent skills

Agent Skills are reusable instruction packages the agent loads on demand. Each skill is a subdirectory holding a SKILL.md in the agentskills.io format: the model sees the skill’s metadata up front and loads its full instructions only when a task calls for it. Strands harness wires skill loading in through the Strands Harness SDK’s skills plugin.

By default Strands harness scans ./.agent/skills for skills. If the directory exists, every skill inside it is registered; if it does not exist, skill loading is a no-op, so nothing happens until you add skills. Point it somewhere else, or at several directories, with skills:

from strands_harness import create_harness
agent = create_harness(skills=["./team-skills", "./project-skills"])

A directory that does not exist is skipped rather than erroring, so a list can include optional locations safely.

A skills directory holds one subdirectory per skill, each with a SKILL.md:

.agent/skills/
release-notes/
SKILL.md
incident-review/
SKILL.md

The SKILL.md carries the skill’s name and description in front matter, followed by the full instructions the agent loads when it decides the skill applies.

Pass None/null (or off on the CLI) to disable skill loading entirely, even if a skills directory is present:

from strands_harness import create_harness
agent = create_harness(skills=None)

For the underlying plugin, see the Strands Harness SDK’s Agent Skills documentation. For the full option list, see the configuration reference.