Platform path (2026-08-08): The legacy AgentDock /
/workflows/newcopilot was removed with F6. The successor is Compose Copilot in the platform chat shell (Create → Describe it, empty chat Build your own workflow,/compose, Discover Build your own / Compose a workflow). Kill switch:PLATFORM_COMPOSE_COPILOT(default on).
Compose Copilot drafts graphs beside chat. On Chat+Nodes (viewport >640px)
the canvas paints the projected proposal while the turn is still in
flight. It stages a change set (graph_proposal) — added / changed /
removed nodes highlighted on Nodes — that you Apply or Discard. The
session is the draft until you Save as workflow (private create or new
version). It never silently rewrites your graph and never starts a billed run
without the cost-confirm card.
What it can do
- Draft a workflow from a description. “A 30-second product ad from a URL: research the site, write a script, generate scenes with review gates, add music, render vertical.” It will ask before guessing when the shape or cost is ambiguous.
- Edit the open workflow. “Add a best-of-3 around the clip generation and keep the highest video-review score.” Proposed changes arrive as a change set — added, removed, and changed nodes highlighted on the canvas — that you apply or discard explicitly. It never silently rewrites your graph.
- Recommend quality patterns. Review gates, retry loops, best-of-N, and approval gates — with the cost impact of each, since every one of those multiplies spend.
- Answer with your real context. It can search Discover for a workflow to reuse, inspect a published graph (compact index, never the full
apiGraph), paste a slice of that graph (selected nodes or class types), list your trained models and their training status, and look up why a past run failed. It cannot list BYOK providers — org BYOK is not on platform. - Recommend a model for the job. Ask “which model for I2V?” and it uses a closed routing table (
recommend_models) — SFW I2V is Hailuo; native H3 only when that flag is on; never an invented or ungranted weight. - Quote and launch runs. Ask “what would this cost?” for an itemized estimate, or “run it with concept X.” A run proposal renders as a card with the inputs filled in and a fresh quote — the run only starts, and credits are only held, when you click Run.
What it deliberately won’t do
- Spend without you. The copilot cannot start a run, hold credits, or publish. Every action with a cost lands as a card you confirm.
- Invent numbers. Prices and quotes come from the same estimator the run button uses — if it hasn’t quoted, it says so.
- Touch provider vaults. Org BYOK is not on platform; the copilot has no customer keys to leak.
Working with it well
- Give the outcome, not the wiring: “consistent character across 6 scenes” prompts it to add subject plates and review gates you might not know exist. You can also train and import custom character LoRAs at
/platform/studio?recipe=characterand bind them with Use character. - Iterate in small steps and apply each change set — smaller diffs are easier to inspect on the canvas.
- Ask “why is this expensive?” — it will break the estimate down by node and suggest where a cheaper tier or a smaller fan-out would do.
Drafts & Save
Compose sessions live in the platform chat session (sidebar). The first brief titles the draft so it is not hidden as an empty Untitled session. Save as workflow creates a private workflow (or a new version when already bound). Publish remains a separate verified-run step.
After a ComfyUI import, the composer is prefilled with unsupported nodes — review and Send for Compose to suggest replacements. Nothing is sent automatically.