Runtime, Connection, and Agent

Component	Lifetime	Responsibility
Agent	declaration	Prompt, model override, tools, subagents, permissions, skills, output schema
actor session	durable	UUID identity, parent/root relationship, FIFO inputs, independent journal, scalar cursor
Runtime	one process	Provider, store, registry, actors, local notifications, live root asks, writer generations
Connection	client attachment	Root event iteration and optional writer authority

Runtime receives the root agent registry. Subagents are discovered transitively from Agent.subagents; actor types are durable, so later activation resolves the same declaration. An optional child display name is stored separately and never changes registry lookup, prompt, model, or permissions.

runtime = Runtime(provider, store, [Researcher], default_model="openai/gpt-5")
root_id = await runtime.create(Researcher, metadata={"tenant": 42})

A root or child actor runs at most one turn at a time. Different actors and independent tool calls may overlap. An idle actor has no permanent worker: a durable input activates a drain task, and the task exits when the inbox is empty.

Only a writable root connection accepts human commands. Opening a newer writer atomically invalidates the previous writer. Readers never own execution, and disconnecting does not cancel accepted work.

Every actor has its own journal. ChildCreated reveals a child UUID. Applications can read durable headers with tree_status() or follow them with watch_tree(), and subscribe to an actor journal only when its detailed events are needed. There is no merged tree stream or composite cursor.

Model guidance layers

Model guidance has four layers:

1. The application behavior prompt comes from Agent.prompt and remains the first system block.
2. The execution environment is one final system block assembled from active capabilities, such as available skill names and child lifecycle semantics. It is omitted when no guidance applies.
3. Tool schemas and descriptions define tool arguments and ordinary usage.
4. Skill bodies provide detailed instructions only after the model loads them on demand with the skill tool.

The execution-environment layer is rebuilt from active framework capabilities for each sample. Compaction may replace only journal-derived history; application prompts, this environment block, and tool schemas are preserved exactly.