Create With AI: Build a Dispatcher Graph with Dynamic Agents, Roles and Prompts
Most workflow builders ask you to design the graph first and think about the team later. Create With AI in Multi-Agent mode does both at once. Describe the outcome and the coordination pattern you need, and CTRL NODE proposes the stages, connections, reusable agents, roles, and instructions as one reviewable draft.
This walkthrough uses the Dispatcher pattern, a useful way to organize systems that receive many kinds of requests. One agent gathers the request, one routing agent decides where it belongs, and a focused specialist handles it. The example is deliberately familiar: a call center that can route a customer to technical support, commercial support, incident management, or a simulated person.
Nothing is created while the proposal is being generated. You review the topology and the team first, then decide whether to turn the draft into a real graph.
Step 1 — Describe the Dispatcher graph
Open Graphs & Workflows → New, choose Create With AI, and switch Agent Mode to Multi-Agent. In this mode, the planner does more than generate task nodes: it can propose reusable agents, write their roles and instructions, and assign each stage to the right agent.
The first stage in our example is a Call Operator. It greets the caller, asks what they need, waits for the answer, and turns that conversation into a structured request. The next stage is the Call Dispatcher, which reads the request and selects the appropriate destination.
AI will propose reusable agents with roles and instructions, then assign workflow stages to them. One agent may be used by multiple stages.
Here is the complete prompt used for the graph:
Create a multi-agent call-center graph that demonstrates the Dispatcher pattern.
Start with a "Call Operator" stage. The operator greets the caller, asks "How can we help you today?", waits for the answer, and writes a clear summary of the request.
Pass that summary to a "Call Dispatcher" stage. The dispatcher classifies the request and routes the full context to the most appropriate destination:
- Technical support for product or technical problems.
- Commercial support for pricing, plans, or sales questions.
- Incident management for outages, urgent failures, or service incidents.
- A simulated person when the request needs a natural human-style response.
Create reusable agents with a name, role, and detailed instructions. Reuse one Domain Specialist agent for the technical, commercial, and incident stages, but give every stage a focused task prompt. After the selected destination responds, return the answer through the Call Operator.
Use a manual trigger. Generate the graph as a draft so I can review the topology, agents, roles, instructions, and stage assignments before creating it.
The important part is not the call-center wording. It is the division of responsibility: intake, dispatch, specialist work, and response. The same shape can route support tickets, documents, leads, security alerts, or internal requests.
Step 2 — Review the AI proposal
Select an agent template and choose Generate Graph. CTRL NODE returns a read-only proposal: nothing has run, no workflow tasks have been created, and the proposed agents have not yet been added to the Team.
For this prompt, the proposal contains seven stages:
- Operator greets and captures request
- Dispatcher classifies and routes
- Technical agent resolves issue
- Commercial agent handles inquiry
- Incident agent logs and responds
- Simulated person responds
- Response
The proposal also creates four reusable agent definitions:
- Call Operator — an intake operator that gathers the request and later returns the answer.
- Call Dispatcher — a router that classifies the request and passes the complete context forward.
- Domain Specialist — one reusable specialist assigned to the technical, commercial, and incident stages.
- Simulated Person Agent — a human-style responder for requests that should be answered as a specific person.
Greet the caller warmly and ask what they need help with. Listen to the response, capture the essential details (intent, key facts, urgency signals), and produce a clear, structured summary of the request without trying to resolve it yourself. Never invent information the caller did not provide; if something is ambiguous, note the ambiguity explicitly so downstream routing can account for it.
Read the incoming request summary and classify it into exactly one of the available categories based on its primary intent. Base the decision on explicit signals in the text (keywords, described problem, nature of the ask) rather than assumptions. State the chosen category and a brief justification, then hand off the full context to the corresponding specialist so they can act without needing to re-ask basic questions.
Review the handed-off request and context, then address it within your domain of expertise: diagnose and resolve technical issues, handle sales and pricing inquiries, or manage and log incidents as appropriate. Ask only for information that is genuinely missing, be precise and actionable in your response, and clearly flag anything that falls outside your domain instead of guessing.
Respond as a specific individual would, in a natural, first-person, conversational tone rather than a generic assistant voice. Stay consistent with the persona's presumed role and knowledge, give direct human-like answers, and acknowledge when something is outside what that person would reasonably know or handle.
This is where Multi-Agent mode becomes more than automatic node placement. The proposal includes a role and detailed operating instructions for every agent. It also understands that an agent is reusable: Domain Specialist appears once in the Team but can be assigned to three different stages, each with its own focused task prompt.
Before creating the graph, review the branch structure, open the full stage prompts, and check every assignment. If the proposal misunderstood the request, return to the prompt or generate another draft. The review step is the control boundary.
Step 3 — Inspect the Designer and generated Team
Choose Create Graph when the proposal matches your intent. The draft becomes a normal, editable CTRL NODE graph. Its nodes and connections appear in Designer, and the proposed agents are created in Team with the generated names, roles, and instructions.
The Designer preserves the planned structure: Call Operator passes the captured request to Call Dispatcher; the Dispatcher exposes the technical, commercial, incident, and simulated-person destinations; and the final response returns through the operator.
The static view below reproduces the completed Run #5 Designer graph and the four agents generated for it. The pre-existing AG1 agent is intentionally excluded.
up to 60sHEARTBEAT: 21:45:30Z
up to 60sHEARTBEAT: 21:45:30Z
up to 60sHEARTBEAT: 21:45:30Z
up to 60sHEARTBEAT: 21:45:30Z
The Team contains only the agents created for this graph. Call Operator is reused for intake and response, while Domain Specialist is reused by the three domain-specific stages.
Generated agent instructions
These are the complete instructions stored on each generated agent.
Call Operator
Role: Intake operator
Model: claude-sonnet-5
Greet the caller warmly and ask what they need help with. Listen to the response, capture the essential details (intent, key facts, urgency signals), and produce a clear, structured summary of the request without trying to resolve it yourself. Never invent information the caller did not provide; if something is ambiguous, note the ambiguity explicitly so downstream routing can account for it.
Call Dispatcher
Role: Router
Model: claude-sonnet-5
Read the incoming request summary and classify it into exactly one of the available categories based on its primary intent. Base the decision on explicit signals in the text (keywords, described problem, nature of the ask) rather than assumptions. State the chosen category and a brief justification, then hand off the full context to the corresponding specialist so they can act without needing to re-ask basic questions.
Domain Specialist
Role: Technical / Commercial / Incident specialist
Model: claude-sonnet-5
Review the handed-off request and context, then address it within your domain of expertise: diagnose and resolve technical issues, handle sales and pricing inquiries, or manage and log incidents as appropriate. Ask only for information that is genuinely missing, be precise and actionable in your response, and clearly flag anything that falls outside your domain instead of guessing.
This single agent is assigned to three graph nodes. The node instruction narrows its work to the technical, commercial, or incident branch.
Simulated Person Agent
Role: Simulated human responder
Model: claude-sonnet-5
Respond as a specific individual would, in a natural, first-person, conversational tone rather than a generic assistant voice. Stay consistent with the persona's presumed role and knowledge, give direct human-like answers, and acknowledge when something is outside what that person would reasonably know or handle.
Generated task instructions by node
Agent instructions define the reusable behavior. Task instructions define the specific responsibility of each node in this graph. The following text comes from the completed Run #5 tasks.
1. Operator greets and captures request
Assigned agent: Call Operator
Act as the call operator: greet the caller, ask 'What can I help you with today?', and wait for their response describing their need. Produce a concise, structured summary of the caller's request including intent, key details, and any urgency or ambiguity signals, ready to be handed to a dispatcher.
2. Dispatcher classifies and routes
Assigned agent: Call Dispatcher
Read the operator's request summary. Classify the request into exactly one category: technical, commercial, incident, or simulated-person, based on its content. State the chosen category with a short justification, and prepare the full context needed for that destination agent to act.
3. Technical agent resolves issue
Assigned agent: Domain Specialist
Handle the request routed as technical: diagnose the described problem using the provided context, propose a clear resolution or troubleshooting steps, and note any missing information needed to fully resolve it.
4. Commercial agent handles inquiry
Assigned agent: Domain Specialist
Handle the request routed as commercial: address pricing, sales, or offer-related questions using the provided context, and give a clear, actionable commercial response.
5. Incident agent logs and responds
Assigned agent: Domain Specialist
Handle the request routed as an incident: log the reported issue with relevant details, assess severity, and provide the caller with next steps or expected resolution timeline.
6. Simulated person responds
Assigned agent: Simulated Person Agent
Handle the request routed to the simulated person: respond in first person as that individual would, using the provided context to give a natural, human-like reply to the caller's request.
7. Response
Assigned agent: Call Operator
The final node uses this generated instruction:
Using the specialists' responses, create a response for the caller.
Each specialist branch connects to this same response task. The operator uses whichever branch completed to prepare the final caller-facing reply.
Step 4 — Pause for caller input, then continue
Execution can pause at the point where a stage needs information that it cannot safely invent. In this run, the Call Operator stops the intake stage and creates a follow-up request for the caller's description of the issue. The graph does not send an incomplete or fabricated request to a specialist.
The caller reports that Windows starts normally, then the mouse and keyboard stop responding around one minute later. They describe the incident as urgent and say that it affects the company's systems. Once the follow-up arrives, the operator turns it into a structured intake result rather than attempting a diagnosis.
“When I start the computer, Windows launches, but after a while the mouse and keyboard stop responding. I think the operating system is frozen. This is urgent: all of the company’s systems do the same thing. Please transfer me to a specialist.”
Status: Captured — ready to route to a specialist; not resolved by the intake operator.
- Intent
- Windows freezes shortly after sign-in; the caller requests an immediate transfer.
- Key facts
- Windows starts normally; mouse and keyboard stop responding about one minute later; the issue is recurring.
- Urgency
- Potential company-wide impact and an explicit urgent escalation request.
The operator does not guess at a diagnosis. It records the missing context and hands a structured, high-priority request to the Dispatcher.
The summarized result captures the intent, observed symptoms, urgency signals, and open questions. Its suggested next action is a high-priority route to the technical specialist. From there, the Dispatcher has enough context to make the routing decision, while the specialist can focus on diagnosis and resolution.
Step 5 — Classify the incident and send the response
The Call Dispatcher receives the structured intake result and classifies it as an incident. The decision is based on the recurring failure, the report of multiple affected systems, and the caller's explicit urgency — not on an assumed root cause.
The Incident Specialist logs INC-2026-08-10-5 and assesses the case as Sev-2 (High) while the company-wide scope is still unverified. It asks for the number of affected machines, peripheral type, recent changes, and the result of a complete restart. If two or more machines are confirmed with the same symptom, the case becomes Sev-1 and is escalated to infrastructure on-call; a single affected machine takes the standard technical-support route.
The final Call Operator response turns that work into a customer-facing message. It asks for the exact machine count, peripheral type, recent updates or policy changes, and the result of a full restart. While scope is confirmed, it recommends making no further changes and testing one machine in Safe Mode. Confirmed multi-machine impact activates a Sev-1 infrastructure response within 30–60 minutes; a single-machine issue moves to technical support within four business hours.
Call Dispatcher
Classification and routing
The recurring one-minute freeze, reported multi-system scope, and explicit urgency make this an incident rather than routine technical support.
Incident Specialist
Log, assess, and guide
Sev-2 (High) while scope is unconfirmed. Escalate to Sev-1 for two or more affected machines; downgrade to Sev-3 for one device.
Call Operator
Final caller response
Acknowledges urgency, gives immediate safe-mode and change-freeze guidance, and explains the 30–60 min / 4 business-hour paths.
We are treating the reported Windows freezes as high priority. Please confirm how many machines are affected, whether the mouse and keyboard are wired or wireless, whether any updates or policy changes were recently deployed, and whether a full restart temporarily restores operation.
This handoff keeps each responsibility narrow: the Dispatcher chooses the route, the specialist assesses and records the incident, and the operator delivers a clear next step without inventing missing facts.
At this point, everything is still under your control. You can reposition nodes, edit task instructions, change an assignment, or refine an agent role before running anything.
One prompt has produced both parts of the system: the workflow that coordinates the work and the reusable team that performs it. Start with Graphs & Workflows → New → Create With AI, select Multi-Agent, and describe the coordination pattern you want to review.