Workflows and automation

A workflow is a saved, reusable plan: a directed acyclic graph of steps with dependencies. Each run keeps its own journal, so you can edit the definition without losing what earlier runs did. Automation can start runs on a schedule or on named events, but only after you approve that exact definition.

#Create a workflow

1. Open Workflows and choose a recipe: feature work, investigation, review, documentation, tests or deployment preparation.
2. Enter a bounded goal and the model and effort, then create the saved DAG. Saving never runs it.
3. Inspect the canvas. Drag to pan, + and − to zoom, 100% for actual size or Fit to see it all.
4. Run workflow shows every step for approval. The run journal then records what actually happened, separately from the definition.

#Write steps by hand

Advanced JSON edits the definition directly. This is the shape a recipe produces:

{
  "name": "Settings refactor",
  "steps": [
    {
      "id": "plan",
      "action": "agent-task",
      "dependsOn": [],
      "timeoutMs": 1800000,
      "input": {
        "engine": "claude",
        "model": "",
        "effort": "high",
        "title": "Planner",
        "prompt": "Plan the settings refactor…"
      }
    },
    {
      "id": "implement",
      "action": "agent-task",
      "dependsOn": ["plan"],
      "input": { "engine": "codex", "title": "Implementer", "prompt": "…" }
    },
    {
      "id": "tests",
      "action": "agent-task",
      "dependsOn": ["plan"],
      "input": { "engine": "codex", "title": "Tester", "prompt": "…" }
    },
    {
      "id": "review",
      "action": "agent-task",
      "dependsOn": ["implement", "tests"],
      "input": { "engine": "claude", "title": "Reviewer", "prompt": "…" }
    }
  ]
}

#How runs execute

• Independent steps run in parallel, and dependents wait for their predecessors.
• Each agent step starts its own session in the project and receives its predecessors' results as context, to verify rather than trust.
• A failed predecessor blocks its dependents. Nothing downstream runs on a broken result.
• Side effects aren't retried. Agent and Design actions aren't retry-safe, so they run once.
• Cancellation is a request. Work that hasn't settled stays uncertain until you check it.
• Interrupted runs aren't replayed after a restart.
• Editing a definition preserves earlier runs and their journals.

#Automation

Workflows → Automation → + Add automation attaches a trigger to a workflow:

New automation starts disabled. Enabling or changing it needs your approval of the exact definition revision and trigger, so editing the workflow later revokes that approval until you approve again.

• Missed and overlapping intervals are skipped, never queued up.
• A failed or uncertain dispatch stops future automation for that workflow until you review it.
• Reusing an event ID returns the receipts it already produced; New event ID requests a distinct delivery.
• Emit an event needs approval too.

Inspect dispatch history next to the run it started. Disable or remove an automation to stop future dispatches.

#Workflows, recipes or teams?

Tasks · Agent teams · Design