Microsoft Agent Framework、Python 用「Agent Skills」を正式リリース

Your Python agents can now pick up reusable bundles of domain expertise (instructions, reference material, and scripts that load only when a task calls for them) through a stable, production-ready API. Agent Skills for Python in Microsoft Agent Framework is stable and shipping: the core skills API has no experimental gate, so you can rely on it in production without the churn of a preview surface. Teams can author skills, release them on their own schedule, and drop them into any agent, backed by the governance controls enterprises expect before agents reach production.

If you’ve been following our earlier posts on file-based skills and authoring modes with script execution, everything described there is now stable and shipping.

What are Agent Skills?

Agent Skills is an open format for packaging domain expertise that agents discover and use on demand. Each skill has metadata and instructions – a SKILL.md file for file-based skills, or equivalent properties in code – optionally accompanied by scripts, reference documents, and other resources. The agent loads only what it needs, when it needs it, keeping the context window lean through a four-stage progressive disclosure pattern: advertise skill names → load instructions → read resources → run scripts.

The result is agents that gain specialized capabilities without bloating their core instructions or their context window – and expertise you author once and reuse across every agent that needs it.

For a full introduction to the format, see Give Your Agents Domain Expertise with Agent Skills.

What you can do with it

Enforce enterprise policy consistently

Package your company’s HR policies, expense rules, or IT security guidelines as skills. An employee-facing agent loads the relevant policy skill when someone asks “Can I expense a co-working space?” and answers from the policy itself – without every policy sitting in context at all times. Because the skill gives every agent the same vetted guidance to follow, employees get consistent, grounded answers.

Turn support playbooks into repeatable workflows

Turn your support team’s troubleshooting guides into skills. When a customer reports an issue, the agent loads the matching playbook and follows the documented steps – so resolution is consistent regardless of which agent instance handles the request.

Compose skills from multiple teams

Teams can build and maintain their skills independently — as file directories in a shared repo, or as packages on your internal PyPI feed — and you assemble them into one agent with no cross-team coordination. The agent picks which skill to use from each skill’s description; there’s no routing logic for you to write.

A great place to find some existing skills is in the Awesome Copilot repo for skills.

Three ways to author skills

The release supports three authoring styles, so each team can pick what fits how they work. All three plug into the same provider, and the agent treats them identically at runtime:

File-based skills – A directory with a SKILL.md, optional scripts, and reference documents. Good for skills that live in a shared repo and are maintained by non-developers or cross-functional teams.

Class-based skills – Python classes that package instructions, resources, and scripts for distribution through normal Python workflows, including internal PyPI packages.

Code-defined skills – Skills created directly in application code. Useful when a skill needs to be generated dynamically or close over application state.

Built for production

Giving an agent new capabilities is only useful if you can govern how it uses them. This release includes the controls you need to run skills in production.

Human-in-the-loop approval. The skills provider exposes three tools the agent calls to work with skills – load_skill (load a skill’s instructions), read_skill_resource (fetch a bundled resource), and run_skill_script (execute a bundled script). All three require approval by default, so nothing loads or executes without oversight – and you can relax it selectively for trusted operations.

Controlled script execution. Class-based and code-defined skill scripts run in-process. File-based scripts are delegated to a runner you provide, so you own sandboxing, resource limits, and audit logging.

Filtering. Expose only a curated subset of a shared skill library to a given agent, with a predicate that can make context-aware decisions based on the requesting agent or tenant.

Caching. Skills are resolved once and reused, with optional per-key isolation so one provider can serve different skill sets to different agents or tenants.

Extensible source pipeline. The underlying source classes are now public, so when the builder doesn’t fit your needs you can compose custom pipelines or integrate skills from your own registries.

As with any dependency that can influence agent behavior, review skill content before deployment, sandbox file-based scripts, and log which skills, resources, and scripts are used.

Getting started

Install the required Python packages, then wire a skills provider into an agent:

import asyncio

import os

from pathlib import Path

from agent_framework import Agent, SkillsProvider

from agent_framework.foundry import FoundryChatClient

from azure.identity import AzureCliCredential

async def main() -> None:

client = FoundryChatClient(

project_endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"],

model=os.environ.get("FOUNDRY_MODEL", "gpt-4o-mini"),

credential=AzureCliCredential(),

)

# Discover file-based skills from a directory of SKILL.md files.

# Skill tools require approval by default. For these trusted skills we opt

# the read-only tools out of approval so the agent can load skills and read

# resources unattended, while running a script still requires approval.

skills_dir = Path(__file__).parent / "skills"

skills_provider = SkillsProvider.from_paths(

skill_paths=str(skills_dir),

disable_load_skill_approval=True,

disable_read_skill_resource_approval=True,

)

async with Agent(

client=client,

instructions="You are a helpful assistant.",

context_providers=[skills_provider],

) as agent:

response = await agent.run("Help me with onboarding.")

print(response.text)

if __name__ == "__main__":

asyncio.run(main())

Why this matters

Agent Skills gives you a standard way to package, distribute, and govern domain expertise for your agents. Teams author skills independently, the builder composes them into a single provider, and approval keeps a human in the loop for anything that matters. With the Python API now stable and shipping, you can build on it in production without riding the churn of an experimental API.

image Agent Skills documentation on Microsoft Learn

image Python samples on GitHub

image GitHub Discussions – share feedback and connect with the community

image Agent Skills Specification

The post Agent Skills for Python Is Now Released appeared first on Microsoft Agent Framework.