Tutorial: your first workflow

In about 20 minutes you will build two agents, connect them into a workflow that routes urgent member requests, run it on a real case, and publish it. Every screenshot below was taken while building exactly this.

At a glance

You will build	Support Triage (an agent that classifies a request), Urgent Reply Drafter (an agent that writes a reply), and Support request routing (a workflow that sends urgent cases to the drafter)
Who	Author to build and run. Approver or Admin to publish
Before you begin	Access to a project; an LLM key in the Vault (ask an Admin)

What you will end up with:

---

Part 1: Build the triage agent

Step 1: Name it

1. In the sidebar, click Agents, then New agent.
2. Under Name, type Support Triage.
3. Under Description, type: Reads a member's support request and returns its category, priority and a one-line summary.
4. Click Next: Behaviour.

The Reference (support_triage) is filled in from the name. Workflows use it to find the agent, so it can never change after the agent is created.

Screenshot: Identity

Step 2: Tell it what to do

In System, paste:

You triage support requests from health-plan members.

Read the request and decide:
- category: one of billing, claims, eligibility, pharmacy, other
- priority: P1 if the member cannot get care today, P2 if it needs an answer this week, otherwise P3
- summary: one sentence a coordinator can read at a glance

Use only what the member wrote. If the request is not a support request, use category "other" and priority P3.

Step 3: Choose a model and its key

1. Click Model. Choose Provider: OpenAI and Model: GPT-4.1 mini.
2. Click Credentials. Under API key, choose your organisation's OpenAI key (for example org/openai_key).

Step 4: Ask for a structured answer

Other steps will route on the priority, so the agent must reply in a fixed shape, not prose.

1. Click Output and switch Structured output (JSON) on. A drawer opens.
2. Leave Example selected and paste a sample answer:

   { "category": "claims", "priority": "P2", "summary": "Member asks why a claim from 12 March was denied." }

3. Click Validate, then Save structure.

Screenshot: Defining the answer's shape from an example

Step 5: Declare its input

1. Click Inputs, then Add input.
2. Name: request · Type: Text · switch Required on.

Screenshot: The request input

Step 6: Create, test and publish

1. Click Create agent (top right).

2. Click the Playground tab. In request, type: Hi, my pharmacy refused to fill my insulin this morning, they said my coverage ended. I only have one dose left. Member ID HP-448120.

3. Click Run. After a few seconds the result shows Passed, with Category: pharmacy and Priority: P1.

   Screenshot: A passing Playground run

4. Click the Publish tab, then Publish v1. (Approver or Admin. If you are an Author, ask one to do this step.)

Why test first? Publishing needs a passing Playground run of exactly this version. It is the proof that what you release works.

---

Part 2: Build the reply agent

Repeat Part 1 with these values. Skip Step 4: this agent replies in plain text.

Field	Value
Name	Urgent Reply Drafter
Description	Drafts a short, calm reply to an urgent member request and says what happens next.
System	You write replies to urgent requests from health-plan members. Write three or four sentences. Acknowledge the problem in the member's own terms, say that a coordinator will call them within two hours, and give one thing they can do right now. Never promise coverage or a decision.
Model / key	OpenAI · GPT-4.1 mini · your OpenAI key
Input	request, Text, Required

Create it, run it once in the Playground, and Publish v1.

---

Part 3: Draw the workflow

Step 1: Start from a template

1. Click Workflows, then New workflow.
2. Click Triage and route.
3. At the top, type the name Support request routing.

Screenshot: The starting shapes

Step 2: Set the Classify step

1. Click the Classify step. Its settings open on the right.
2. Click Choose an agent and pick Support Triage. The step takes the agent's name.
3. Under Inputs, for request, choose From the case. The case field is filled in as request.

Screenshot: The agent step's settings

"From the case" means: ask whoever starts the run for this value.

Step 3: Set the routing rule

1. Click the Route step.
2. On the Urgent path, click Add a rule to say when this path is taken.
3. Click Field and choose Support Triage → priority.
4. Leave the operator as is and type P1 as the value.

Screenshot: Choosing the field to route on

Anything that is not P1 takes the Routine path (the Otherwise path).

Step 4: Set the urgent agent

1. Click Handle urgent, then Choose an agent → Urgent Reply Drafter.
2. Under Inputs → request, choose From the case.

Step 5: Save

Click Save. The badge beside the name reads v1 draft · not live.

Screenshot: The finished map

---

Part 4: Run it

1. Click Run. A window asks for the fields the steps read from the case.
2. Paste the insulin request into request and click Run.

Screenshot: Starting a run

Watch the steps light up. The run takes the Urgent path; Routine is greyed as Skipped. The bar reads Completed · Ended at Done · completed · 6/6 steps.

Screenshot: The finished run

Click any step in the run panel to see what it received and returned.

---

Part 5: Publish it

1. Click Publish v1. The dialog confirms "A test run of exactly this version finished."
2. Click Publish v1.

Screenshot: Publishing the workflow

The badge now reads v1 live.

---

What you learned

You did	The idea behind it
Wrote instructions and chose a model	An agent is a declaration, not code
Asked for a structured answer	Later steps can read named fields like priority
Tested before publishing	Only tested versions can be released
Mapped inputs "From the case"	Each step says where every value comes from
Routed on a field	A Condition chooses one path; the others are skipped
Published	Released versions are fixed; editing makes a new draft

Next steps

• Give an agent a tool → Tools
• Let it answer from your documents → Knowledge
• See the run's full record → Traces
• Let another system call an agent → Published agents