Tool

A tool is a named capability the harness offers the model: read this file, run this command, search for this pattern, fetch this URL. Each one comes with a description and a parameter schema, and the whole list travels with every model provider request. When the model wants to use one, it writes a tool call as text; the harness runs the underlying code and hands back a tool result. Nothing runs inside the model. Execution is the harness's job.

The tool list is the agent's reach. A shell tool alone covers most of what a developer does at a keyboard, which is why so much agent work flows through it, and why it deserves the tightest permissions. Take a tool away and the agent doesn't complain, it improvises: with no browser tool it will guess what a page looks like, with no test runner it will 'reason about' whether tests pass. Missing tools produce fabrication, not errors.

Tools have a standing cost too. Every definition takes up context on every request, and a long list of overlapping tools makes the model pick the wrong one more often. Add tools deliberately: one per capability the task genuinely needs, with a description that says when to use it. MCP is how you plug in tools the harness doesn't ship with.