Capabilities
Coming soonA Harness defines what an agent can use. Start with the capabilities your task needs, then configure their permissions and runtime requirements.
| Task | Guide |
|---|---|
| Call typed Python functions | Tools |
| Use tools from external systems | Connectors & MCP |
| Run code and work with files | Sandbox |
| Load reusable procedures on demand | Skills |
| Gate tool calls and request approval | Permission |
| Check structured-output support and limitations | Structured output |
Before you enable a capability
Before you enable a capability
Read the selected guide before adding it to your harness. Capabilities have different requirements: connectors can require a hardened worker, sandboxes have different isolation properties, and skills are early in development. The structured-output guide documents a known limitation with response_format.
To customize execution around model and tool calls, see Custom hooks. For a complete agent using tools, follow the quickstart.