Start here
Core concepts
Agents, spawning, messaging, budgets, the step journal.
| Concept | What it means |
|---|---|
| Agent | An actor with a role, a goal, capabilities (which decide its tools), a budget and a status (Thinking, Executing, Waiting, Completed…). Status changes are explicit transitions, and each one emits an event. |
| Spawning | An agent calls spawn_agent. The runtime checks depth, children per agent, team size, team-shape policy and budget, then creates the child. A child's tools and permissions can only narrow. |
| Messaging | Agents find each other with find_agents and talk with send_message, without relaying through the root. Messages are untrusted input and are persisted. |
| Budgets | Tokens, tool calls, dollars and time. A child's budget is carved from what its parent has left and reserved against it, so a whole tree never spends more than its root was given. |
| Completion | complete_task with a summary, artifacts, evidence and remaining work. The root decides when the whole goal is done. |
| Durability | Each step of an agent's turn is journaled before the next runs. After a crash, work resumes. A non-idempotent tool call that was cut off is reported as "outcome unknown", never repeated blindly. |
| Organizations | Everything belongs to one organization (a tenant). Another organization's agents, data and runs don't exist as far as you can see. Roles are Viewer, Member, Admin and Owner. |