Agent Skills
Agent Skills are portable folders that package task-specific instructions and, optionally, scripts and supporting resources for an AI agent. A SKILL.md file supplies the skill's name, description, and operating instructions; additional files can be loaded or executed when the task requires them. Progressive disclosure lets the agent discover a compact catalog first and bring detailed material into context later. A skill is therefore a reusable context and procedure bundle, not a model capability guarantee or an external service connection by itself.
Origin and context
Anthropic announced Agent Skills on 16 October 2025 and described a filesystem-based format used across several Claude products. Its engineering explanation emphasized composability and progressive disclosure: metadata supports discovery, the main instructions load when relevant, and linked files are accessed only as needed. A dated Microsoft Agent Framework engineering article subsequently documented file-based skills with SKILL.md, references, and scripts alongside other authoring modes. That independent implementation supports treating Agent Skills as a cross-platform practitioner concept while retaining Anthropic in the alias because the base catalog used the vendor-qualified name.
Why it matters
Skills separate reusable task knowledge from a single conversation or a growing system prompt. Teams can version domain procedures, ship examples or deterministic utilities with them, and let an agent select only the material relevant to the current request. This can make agent behavior easier to maintain and share, but it also turns skill folders into a software and content supply chain. Writers need to define triggers and boundaries clearly; operators need to review scripts, dependencies, permissions, and provenance before allowing execution.
Example
A document-production skill contains a SKILL.md file that explains when to use it, a template, a rendering script, and a checklist. The agent sees the skill name and short description during discovery. When asked to create the document, it loads the full instructions and opens the template; it runs the script only if rendering is required. A paragraph pasted permanently into a system prompt is reusable guidance, but it is not an Agent Skill package in this filesystem-based sense.
How it differs
Model Context Protocol
Model Context Protocol standardizes how an AI application connects to external tools and data sources. Agent Skills package procedures and local resources that tell an agent how to perform a class of work. A skill can explain how to use an MCP tool, but installing the skill does not create the connection or grant access.
Agent harness
An agent harness is the runtime that manages model calls, tools, state, permissions, and execution. Agent Skills are artifacts that such a runtime may discover and load. The package does not replace the harness's approval, sandboxing, or observability controls.
Maturity and evidence
Maturity is rated 3. The concept has dated primary documentation, a concrete file convention, and an organizationally independent framework implementation using the same package structure. It remains short of maturity 4 because the reviewed evidence does not establish a neutral standards body, broad long-term compatibility guarantees, or adoption across many independent runtimes. Validation and trust behavior can still differ between implementations even when the folders look similar.
Limits and open questions
A well-written SKILL.md file does not prove that the agent will select the skill correctly, follow every instruction, or produce a valid result. Packages may include executable code or request access to sensitive systems. Anthropic advises installing skills only from trusted sources and auditing their contents. Microsoft's implementation makes skill-tool approval the default and recommends production script runners with sandboxing, resource limits, input validation, and audit logging. Compatibility still needs testing because a package can rely on tools, paths, libraries, or runtime behavior unavailable elsewhere.
Related terms
References
- Equipping agents for the real world with Agent SkillsAnthropic · 2025-10-16 · class A
- Agent Skills in .NET: Three Ways to Author, One Provider to Run ThemMicrosoft Agent Framework · 2026-04-13 · class A
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as context engineering skill.
This term is also covered in the Skills Atlas as workflow orchestration skill.
This term is also covered in the Skills Atlas as human in the loop ai skill.