Skip to main content
Guide: MAF workflows. --wait prompts for HITL inline.

workflow

Create from one instruction

--instruction sends the whole sentence to the MAF planner, the same one the dashboard uses. The planner splits it into steps and picks an agent for each step. --name and --description, when given, replace the planner’s own. --instruction can’t be combined with --step or --agent.

Drafts

--draft on create saves the workflow as a draft instead of deploying it live. Review it with workflow get, change it with workflow update, then deploy it with workflow promote.

Run-time content

run --content "<text>" folds data into step 0’s task for this execution only, before planning. Re-run the same saved workflow against different input each time:

execution

Stop executions

Pass one or more execution UUIDs, or --workflow <name-or-uuid> to stop every unfinished run of that workflow. You can’t pass both.
  • A queued run never starts.
  • A running run is cut short where it is. Its spend so far is still recorded.
  • A paused run’s open question is withdrawn.
Runs that already finished are left as they are and listed under Already finished. Only runs of your own workflows can be stopped. If none of the IDs you gave match one of your runs, the command fails. --json prints the stop report: stopped, already_finished, and not_found.

trace

trace runs one instruction end to end and reports every stage. It does in one command what workflow create --instruction, workflow run --wait, and execution get do in three:
  1. Sends the sentence to the MAF planner and shows how it was split and which agent each step was given.
  2. Runs the workflow, reporting each step as it finishes.
  3. Prints the result with its token breakdown, and the commands to re-run or inspect it.