Quickstart: orchestrate your first workflow
From an empty Canvas to a validated, simulated and exported multi-agent workflow in about ten minutes.
Updated September 25, 2026 · 3 min read
Before you start
All you need is a browser and an account. Simulation runs cost nothing and need no API key; for Live runs you'll add your own Claude API key later (see Integrations and the Vault).
The five building blocks
| Node | What it is | Can link to |
|---|---|---|
| ⚙️ Root Orchestrator | The entry point: the global behavior rules, the analytical engine and the token budget. One per workflow. | Sub-Agents, Skills, References |
| 🤖 Sub-Agent | A specialist the Root delegates to, with its own role and behavior rules. | Sub-Agents, Skills, References, Connectors |
| 🧩 Skill | A tool: a function name, a description and a JSON Schema for its input. | Connectors, References |
| 📚 Reference | Context to read: a document, a URL or data, with reading guidelines. | Nothing (passive) |
| 🔌 Connector | Access to an external tool (an MCP server) through a credential kept in the Vault. | Nothing (passive) |
Build it step by step
Open the Canvas
Go to the Canvas. A demo workflow is already there; Reset demo in the top bar brings it back at any time.
Set up the Root Orchestrator
Select the ⚙️ Root. In the inspector, write the global behavior rules: who the system serves, what it must never do, when to hand over to a person. Keep the target on Claude Code.
Add a Sub-Agent
Drag 🤖 Sub-Agent from the Assets panel onto the Canvas (or click it), then drag from the Root's bottom handle to the Sub-Agent's top handle. Give it a role and its behavior rules. You can also say “Add a Sub-Agent called Researcher”: see Voice commands.
Give it a Skill
Add a 🧩 Skill under the Sub-Agent. Name the function (
search_invoices), describe when to use it, and define its input in the JSON Schema field.Add context and access
A 📚 Reference gives the agent something to read; a 🔌 Connector gives a Skill access to a tool such as Stripe or GitHub. The Connector Hub (Assets panel) adds them pre-configured.
Check the Pre-Flight
The Pre-Flight panel (bottom right) lists every problem: missing fields, orphan nodes, invalid schemas. Click an item to jump to its node. Aim for “All nodes valid and connected”.
Simulate a run
Press ▶ Run workflow. In Simulation mode the console streams scripted events: the Root thinks, delegates, the Sub-Agent calls its Skill, and each node lights up on the Canvas while it works. Free and unlimited.
Export the Blueprint
Export Blueprint downloads a ZIP for Claude Code:
CLAUDE.mdwith the orchestration, one file per Skill and.mcp.jsonwith your Connectors. Secrets appear only as${NAME}placeholders.
Rules the Canvas enforces
The graph flows top-down from the Root. Links that make no sense are refused, with the reason:
- One Root Orchestrator per workflow, and nothing links into it.
- References and Connectors are passive: they never have children.
- A Skill can't call another Skill or a Sub-Agent; delegating is an agent's job.
- No cycles and no duplicate links.
Next steps
- Prompt Engineering on the Canvas: prompts that route well.
- Integrations and the Vault: connect real tools and run Live.
- Get the most out of the Free plan.