Where Do AI Agents Discover Tools - The Skills System Explained
One of the most underappreciated problems in AI agents is tool discovery. How does an agent know which tools it has access to, and when should it use them? The answer is evolving fast, and SKILL.md files are at the center of it.
The Tool Discovery Problem
When you give an AI agent access to dozens of tools - file operations, browser automation, API calls, database queries - it needs a way to know what is available and when each tool is appropriate. Dumping every tool definition into the system prompt wastes context and confuses the model.
SKILL.md Files as Tool Manifests
A SKILL.md file describes what a tool does, when to use it, and how to invoke it. Think of it as a README specifically for AI consumption. The agent reads these files to understand its capabilities without needing every tool loaded into memory at once.
This pattern is powerful because it is composable. Drop a new SKILL.md into a directory and the agent can discover it automatically. Remove it and the capability disappears cleanly.
Dynamic Discovery vs Static Configuration
Static tool lists require manual updates every time you add a capability. Dynamic discovery - where the agent scans for available skills at runtime - scales much better. The agent checks a registry, finds relevant skills for the current task, and loads only what it needs.
This is similar to how MCP servers work, but at a higher abstraction level. Instead of individual tool definitions, you get packaged capabilities with documentation, examples, and usage guidelines built in.
Why This Matters
Better tool discovery means agents that are more capable without being more confused. They use the right tool for the job instead of defaulting to the most general one. And you can extend their capabilities just by adding a file.
Fazm is an open source macOS AI agent. Open source on GitHub.