Configure an Agent
Save reusable instructions, model, tools and Environment configuration.
Open an Agent to configure it alongside the live Preview pane.
Identity
Name and description identify the optional Agent preset in lists and delivery surfaces. Execution state belongs to each Session; there is no Agent type to select.
Runtime and model
Choose a runtime first, then select a model available through its resolved Provider key. If the model list is empty or a runtime shows Needs key, configure the matching Provider in the active Project.
For Pi, explicitly select the runtime and a compatible model from your Project. Built-in models use their catalog protocol; custom models use the protocol selected in the provider credential.
System prompt
Describe the Agent's role, boundaries, answer style, and when it should ask for clarification. Prefer explicit operating rules over vague personality prose. Test both normal and failure cases before publishing.
Skills
Attach reusable instruction packages from the active Project. Skills are loaded for new sessions and read by the Agent when the task calls for them. A missing Skill attachment is reported rather than borrowed from another Project.
MCP servers
Attach authorized Remote HTTPS MCP connections from the active Project. A connection can be attached before authorization, but its tools are usable only when the connection is enabled and authorized.
Environment
Choose the reusable runtime template for packages, setup script, variables and network policy. Without an explicit Environment, the Project default applies. Each new Session freezes the selected revision. Both Full and Limited network policies apply to Session-owned execution. See Environment network policies.
Save and publish behavior
Saved edits apply to new Sessions using the preset. Existing Sessions keep their admitted configuration. Publishing creates a live version for v1 and other published delivery flows.