Agent model

Understand managers, specialists, subagents, sessions, tools, and runtime boundaries.

Agents

An agent is a configured runtime process with a prompt, model provider, tool set, memory settings, and optional schedules.

The daemon stores the configuration and launches agents through Docker so each agent runs in an isolated container.

Subagents

A parent agent can expose another agent as a callable tool. This is how manager-style workflows delegate focused tasks to specialists.

Each subagent has its own prompt, model, tools, and memory.

Sessions

A session holds the full state of one piece of work — the conversation, tool results, artifacts, and any subagent sessions it spawned. Splitting long-running work into focused sessions makes it easier to review what happened and pick up where you left off.