A dynamic workflow turns a goal into a multi-phase run you can watch, restart in part, and save for reuse. Ask for one in plain language and the agent authors the orchestration script itself, then runs it in the background — you get a run id immediately and the result arrives back in the chat when it finishes.
Reach for a workflow instead of a plain sub-agent fan-out when the work has distinct phases, when you want to see it progress, or when you expect to re-run part of it after changing your mind about one step.
Say what you want. "Use a workflow to compare these three approaches and recommend one" is enough — the agent passes your goal as the intent, and the authoring and launching happen in one step. You do not write the script.
Two watching surfaces show a live run:
- The chat side panel's Workflows tab, next to Changes and Subagents.
- Customize → Workflows, which is the saved library rather than the live view: it creates, edits, and runs reusable workflows.
A run streams to the panel while it executes and injects its result into the chat on completion, so you do not have to poll it yourself.
| Ask for | What you get |
|---|---|
| The live status | Whether the run is running, finished, failed, or cancelled, plus how many agents and events it has produced |
| The full result | Every phase, each agent's outcome, the logs, and the final return value |
| Recent runs | The newest runs first, with their status |
| Cancel | The run stops |
A run that ended without a usable return value still reports the agent payloads that did complete, plus a per-agent failure reason for each one that did not — so a partial failure is diagnosable rather than a bare error.
This is the reason to prefer a workflow over a one-shot fan-out. A restart replays the unchanged prefix from cache and re-executes only from the step you name: agent calls before that point reuse the prior run's results without spending a model call, and calls from that point on run fresh. Restarting from the very beginning re-runs everything.
Each restart produces a new run id, so the original run's record is preserved rather than overwritten.
A workflow saved to the library can be run again by name instead of re-authored from intent. Ask what is already saved before describing a new workflow — an existing one that fits skips the authoring step entirely.
In dashboard chat, /workflow lists the saved workflows, and /workflow <name> [input]
runs one by its name (the slug the list shows), passing the rest of the line as
its input. The run starts at once and its result appears in the chat when it
finishes.
Credentials and exfiltration-shaped URLs are stripped from every workflow response before it reaches you, including from mapping keys and not only values — agent output is parsed into these structures, so a credential can arrive as a key.
Start a workflow from the member's chat to keep its workers in that member's memory. Saved workflows and restarted steps keep the same memory boundary. Changing a worker's role does not switch its memory. Other members cannot read or control that run through their tools.
If the member's memory is missing, damaged or archived, the run stops rather than switching to Global memory. Restore the member's memory before retrying. Private execution needs the supported Linux/WSL namespace sandbox or macOS outer Seatbelt sandbox; turning the sandbox off is not a recovery path.
- Subagents: parallel fan-out when the work needs no phases or restarts
- Task runner: autonomous multi-step execution from a spec file
- Dashboard: the chat side panel and the Capabilities tabs