Skip to main content
A tool owns one callable capability. It defines input, output, and execution. Tools exist because agents and workflows need safe ways to act. The model can choose a tool, but the tool owns the deterministic code that runs.

​
Characteristics

  • Typed input describes what the caller must provide.
  • Execution performs one operation.
  • Output returns a structured result the caller can use.
  • Errors describe why the operation could not complete.

​
Boundary

Tools can be local project files, remote integration tools, or MCP-exposed capabilities. The caller chooses when to invoke them. The tool owns how the work runs. Keep tools focused. A tool should do one thing, validate its input, and return a clear result. If the operation grows into multiple stages, approvals, or retries, move the coordination into a workflow.

​
Wrong fit

Do not use a tool as a hidden workflow, long-running run, or large integration layer. Use a workflow for process, a run for durable execution, and an integration for reusable external service access. For implementation steps, see Tools.