Introduced dynamic, discoverable skill packages that agents load per-task instead of bundling all capabilities upfront.
Engineering BlogIntroduced Agent Skills — organized folders of instructions, scripts, and resources that agents discover and load dynamically. Published as open standard for cross-platform portability (Dec 2025). Complement MCP servers by teaching agents complex workflows.
An architecture where agents dynamically discover and load specialized knowledge packages (skills) tailored to specific tasks, rather than loading all available tools upfront. Skills contain instructions, scripts, and resources bundled together. This enables emergent specialization — the agent becomes an expert in relevant domains without needing to load knowledge irrelevant to the current task.
Skills are designed as reusable, combinable building blocks. Agents can compose multiple skills together to handle complex workflows that require expertise from different domains. The skill ecosystem enables cross-domain problem-solving where an agent can leverage knowledge from multiple specialized skills within a single task.
The process agents use to identify which skills they need for a given task. The agent analyzes the problem, recognizes required expertise areas, and loads only those skills. This reduces context overhead while ensuring all necessary knowledge is available — a key advantage over fixed tool loading.
The mechanism enabling agents to find available skills and choose appropriate ones dynamically. Quality control at scale becomes critical here — as the skill ecosystem grows, mechanisms for discovering high-quality skills and preventing low-quality or malicious skills from being loaded become essential infrastructure challenges.