Skills

Agent Skills are bundles of markdown and scripts that give the agent domain knowledge on demand. Each skill follows the open AgentSkills.io standard: a SKILL.md file with YAML frontmatter (name, description) and markdown instructions, plus optional scripts/, references/, and assets/ directories.

Skills use progressive disclosure: metadata is injected into the system prompt upfront (~100 tokens), and full instructions are loaded on demand via a tool call. This avoids flooding the context window with instructions the agent may not need.

The harness supports four skill sources:

Skills are fetched once per session on the first invocation. Within a session, skills persist on disk across multiple invocations. When the VM expires and a new session starts, skills are re-fetched to guarantee freshness.

When you use a custom container image with Git, Amazon S3, or AWS skills, include tar in the image. Give the container user write access to /tmp. For details, see Custom environment (container images).

You can set skills as a default on the harness (via CreateHarness or UpdateHarness), or override per invocation. Invoke-time skills are appended after create-time skills; if both define a skill with the same name, the invoke-time version wins.

AWS Skills

AWS skills are pre-built skills that enable your agents to interact with AWS services. They are organized hierarchically and selected via glob patterns. View their source on GitHub.

Enable all AWS skills

Enable skills by category

Use glob patterns to enable specific skill categories:

Enable a single specific skill

response = client.invoke_harness(
    harnessArn=HARNESS_ARN,
    runtimeSessionId=SESSION_ID,
    skills=[{"awsSkills": {"paths": ["core-skills/aws-cdk"]}}],
    messages=[{"role": "user", "content": [{"text": "Create a CDK stack for a Lambda function."}]}],
)

Combine multiple patterns

skills=[{"awsSkills": {"paths": ["core-skills/aws-cdk", "core-skills/aws-serverless", "specialized-skills/storage-skills/*"]}}]

Git (HTTPS) skills

Clone a skill from any public or private Git repository. Supports subdirectories within monorepos using sparse checkout.

Git fetch must complete within 60 seconds. If the repository requires internet egress, ensure your VPC has a NAT gateway (same requirement as remote MCP servers and custom container pulls).

Amazon S3 skills

Fetch a skill from a customer-owned S3 bucket. Uses the harness execution role credentials.

Filesystem path skills

Reference a skill already on the harness filesystem - baked into the container image or installed at session start via InvokeAgentRuntimeCommand.

Bake into the container image

Include the skill directory in your custom image:

COPY skills/xlsx .agents/skills/xlsx

Install at session start

Use InvokeAgentRuntimeCommand before the first agent invocation. The AgentCore CLI wraps it in agentcore harness exec. Pass the harness ID, which is the part of the harness ARN after harness/. agentcore status shows the harness ARN. The session ID must be 33 to 100 characters. Use the same session ID when you invoke the harness. The harness must already list the path skill, as shown in Filesystem path skills.

SESSION_ID="$(uuidgen)"

agentcore harness exec \
  --id "<harness-id>" \
  --session-id "$SESSION_ID" \
  --command "git clone --depth 1 https://github.com/anthropics/skills /tmp/skills && cp -r /tmp/skills/skills/xlsx .agents/skills/xlsx"

agentcore invoke \
  --harness my_harness \
  --session-id "$SESSION_ID" \
  --prompt "Find errors in the Excel files."

Combine multiple skill sources

All four source types can coexist in a single payload:

response = client.invoke_harness(
    harnessArn=HARNESS_ARN,
    runtimeSessionId=SESSION_ID,
    skills=[
        {"awsSkills": {"paths": ["core-skills/aws-cdk"]}},
        {"git": {"url": "https://github.com/anthropics/skills", "path": "skills/docx"}},
        {"s3": {"uri": "s3://my-bucket/skills/company-style/"}},
        {"path": ".agents/skills/xlsx"},
    ],
    messages=[{"role": "user", "content": [{"text": "Help me with this project."}]}],
)

Error handling

All fetch failures fail the invocation with a descriptive error. Skills are never silently skipped.

Related topics