Orchestration

Orchestration turns one natural-language request into a real orchestrator run that can create and control normal Agentree child runs.

Agentree no longer uses executable YAML flows as the product surface. Saved automation is expressed as prompt templates that the orchestrator reads, previews, and executes through the same run lifecycle as normal agents.

Prompt Templates

Use @node references to describe reusable capabilities:

Create three parallel @planner runs.
Create a @synthesizer sibling that compares their results.
Create a @reviewer sibling for the synthesized result.
If review finds issues, create a @fixer child from the synthesizer.

Node templates are prompts with lightweight metadata such as default agent/model hints. They are not YAML programs, and they do not contain loops, branches, expressions, or output contracts.

You can also write one-off instructions:

Create a node that checks whether the migration plan includes restart recovery tests.

Preview And Confirmation

Complex orchestration prompts preview before child runs start. The preview shows the planned children, parent placement, and wait groups:

Will create:
- 3 x @planner as siblings
- 1 x @synthesizer as sibling
- 1 x @reviewer as sibling
- 1 x @fixer as child of synthesizer, if needed

Confirm the preview to start the child runs. Canceling the preview creates no child runs.

Lifecycle

An orchestrator is a normal persisted assistant run. Child runs are also normal Agentree runs with an additional controlled_by relationship pointing back to the orchestrator.

That means child runs keep normal graph history, commits, status, cancellation, and recovery behavior. The orchestrator records durable orchestration events for preview, confirmation, child creation, waits, retries, cancellation, and completion.

Graph

The graph renders orchestrators as real nodes. Orchestrator-created children remain normal graph nodes, and orchestrator hovercards summarize controlled child runs.

Use orchestration when a task benefits from planning, multiple independent attempts, review, or follow-up child runs. Use a normal send when you want one focused agent run.