Introducing MadoAgent: Orchestrate Your Canvas in Plain English
A persistent panel that turns 'spawn an agent, wire it up, hand it a task' into a sentence instead of a sequence of clicks — opened with Cmd+Shift+M.
Today we're shipping MadoAgent: a persistent panel that turns "spawn an agent, wire it up, hand it a task" into a sentence instead of a sequence of clicks. Instead of a fixed menu of canned actions, MadoAgent is a real tool-calling loop wired to dozens of orchestration actions that act directly on your live workspace — spawning tiles, drawing connections, dispatching workers, recalling memory, and more. When you ask it to do something, it actually does it on the canvas.
What you can do now
- Describe orchestration in plain English — "start a Claude agent in my project," "connect these two agents for peer review," "dispatch a Codex worker for this refactor." MadoAgent turns each sentence into the corresponding canvas actions.
- Spawn, connect, and assign work to agents without manually placing tiles and drawing connections.
- Dispatch headless Codex workers for self-contained tasks via Orchestrator Mode (see Orchestrator Mode).
- Run multi-fork explorations of a problem with Ensemble Exploration (see Ensemble Exploration).
- Recall and remember context from your memory store, and operate the terminal autonomously for goal-driven work.
How to try it
- Press Cmd+Shift+M to open the MadoAgent panel from its pill in the top-right of the canvas.
- Type what you want in plain English — for example, "start a Claude agent in my project."
- Watch it happen on the canvas instead of assembling it yourself.
- Try one of the suggested prompts the onboarding tour offers, like "Connect these two agents for peer review" or "Dispatch a Codex worker for this refactor."
Three levels of orchestration
MadoHub now has a clear answer to "how do I wire agents together," in three tiers from fastest to most flexible:
| Level | Tool | Best for |
|---|---|---|
| Templates | Flows | An off-the-shelf pattern — peer review, pipeline, watcher — dropped in one click |
| Custom orchestration | MadoAgent (⌘⇧M) | Anything that isn't a stock topology: specific agents, specific tasks, ad hoc wiring |
| Manual | Drag between tiles | A one-off connection you want full control over, no template or conversation needed |
None of the three replaces the others — pick whichever matches how deliberate you want to be about the result.
The command palette moved
Landing alongside MadoAgent: a keybinding rebind that's been a long time coming. Cmd+K now opens a global search overlay — full-text search across sessions, messages, tool calls, activity, and terminal scrollback, for finding something that already happened. The command palette, for running an action right now, moved to Cmd+Shift+P. The two are mutually exclusive — opening one closes the other — so you're never stuck with both stacked on screen. If Cmd+K still pulls up the palette in your muscle memory, that's the thing to relearn.
Why it matters / use case
MadoHub's power is wiring multiple agents together, but the mechanics of doing it — placing tiles, drawing connections, configuring triggers — add up when you do it often. MadoAgent collapses those mechanics into a sentence. For a one-off custom topology that no Flow covers, you describe it and it appears. For repetitive setup, Flows stay faster; for precise control over a single connection, manual wiring is still there. MadoAgent fills the middle: anything bespoke enough that a template doesn't fit but common enough that you don't want to build it by hand every time.
Related docs
What's next / limitations
MadoAgent's orchestration actions act on your live workspace but are scoped to setup and routing — it won't rewrite your codebase or make changes outside the canvas without you directing it to. For the full list of what it can do, see the MadoAgent docs.