Agents and Runtimes
In brief
An Agent is the runtime target for a session. Its specification declares the behavior and capabilities that a worker resolves when it performs a turn.
Ownership
An Agent can declare model policy, prompt and feature configuration, capability policy, MCP references, A2A connections, and a runtime kind. Native agent execution and ACP-backed execution have different runtime requirements, but both run through Talon’s durable session model.
Lifecycle
The gateway stores the Agent resource. When a session needs work, the worker resolves the agent’s effective configuration and allowed capabilities for that execution. Changes to an Agent affect later work; a persisted session remains the record of work that already occurred.
Relationships
Agents are not derived from Templates. Templates and Deployments are declarative resource fan-out mechanisms. Agents use Skills and Tools, MCP, and Sandboxes to acquire runtime capability.