Back to docs

Orchestrator & workflows

The orchestrator is a code-aware line editor for composing agent workflows. You define steps, grant tools, apply skills, and wire transitions — then run the whole thing against your provider with a live progress card in the chat.

Line format syntax

Workflows are composed as structured text with a handful of line types:

# Research brief
## @step:gather
@tool:memory
@tool:time
@skill:research
→ @goto:summarize
## @step:summarize
@tool:todo

# Title

The workflow title.

## @step:name

Starts a step definition.

@tool:name

Grants a tool to the current step.

@skill:name

Applies a skill hint to the current step.

→ @goto:name

Adds a transition to another step.

Editing & autocomplete

The editor renders the workflow with live syntax-aware coloring and per-line editing: Enter to split a line, Backspace to merge, Tab to indent. As you type, autocomplete offers completions for @step:, @tool:, @skill:, and @goto: tokens. An "ask AI to modify this harness" prompt box lets you edit the workflow in plain language.

Converting to & from YAML

The line format converts bi-directionally to strict YAML (title + steps[]):

Line format
# Research brief
## @step:gather
@tool:memory
→ @goto:summarize
YAML
title: Research brief
steps:
  - name : gather
    tools:
      - memory
    gotos:
      - goto: summarize

The current orchestration is saved onto the active agent and the active harness, so it is always ready to run or schedule.

Running workflows

Workflows actually execute against your provider. Each step runs the agent loop with only that step's granted tools, then transitions via @goto. A live agentRun progress card in the chat shows step-by-step progress, per-step assistant messages, and tool-run details.

Per-step tool grants

Each step only sees the tools you granted it (schema filtering) — no more, no less.

@goto conditions

Transitions can route on success, error, contains:, or match: conditions.

Loop guard

A loop guard prevents infinite @goto cycles from running forever.

Live agentRun card

A real-time progress card in the chat tracks each step as it runs.

Ways to trigger a workflow

A workflow can be kicked off from four places:

Editor Run button

Hit Run in the orchestrator editor to execute the workflow you're working on.

Hub quick action

"Run Active Agent" is one of the hub's quick actions.

Harness play button

Every harness row in the sessions drawer has a play button to run it directly.

Scheduler workflow triggers

A scheduled task can pick a harness; when the cron fires, it executes the harness's steps instead of a bare prompt.