Definitions and executions
An agent definition belongs to a project. A run executes that definition with a particular input and runtime target. Editing a definition and starting a run are separate operations; the Execution API manages runs and their progress. Definitions stored in project source can have different versions on different branches or in deployed environments. The selected source determines which definition a read or edit addresses.Capabilities
Prompts provide reusable instructions. Tools expose callable actions, skills package reusable behavior, and resource documents provide project configuration. An agent’s capabilities describe what is available to that agent. Tools from connected services also depend on the Integrations API and the caller’s access to a connection. Attaching a tool to an agent does not grant access to the external account. A support agent might combine triage instructions, a ticket-reading tool, and knowledge lookup. The definition describes those capabilities; each incoming question produces a separate run with its own input and result.Discovery and installation
The accessible-agent collection lists agents across projects you can access. The template catalog describes agents that can be forked or installed. A public catalog entry must be installed or forked into a project before it can run. Use the returnedactions to determine what the caller can do.
Project lists identify runnable definitions with
source_project_agent or installed_project_agent. Check can_run and runtime_readiness; readiness does not replace authorization.
Use a run’s run_snapshot to identify the agent definition, runtime target, and requester selected for that execution.
The Evaluations API tests agent behavior against datasets and metrics.