Lab 1 · Configure your first agent

Objective

Build a working WISMO agent from scratch: pick the use case, define the SOP, wire up a tool, run a basic eval, and see it respond to a customer query.

Time: 60-90 minutes Prerequisites: Access to the AgentFleet Dashboard (demo environment). See Cowork & Claude tools for setup.

The scenario

You’re deploying a WISMO agent for a fictional B2C LSP called “QuickShip.” They handle 5,000 deliveries/day and get ~2,000 WISMO calls daily. You need an agent that can:

1. Authenticate the customer (by order number or phone)
2. Look up the shipment status
3. Tell the customer where their order is and when it’ll arrive
4. Escalate if the shipment has an exception

Step 1: Define the SOP (15 min)

Before touching any tool, write out the Standard Operating Procedure. This is what a human agent would follow — now you’re teaching it to an AI.

Your SOP should cover:

• How to greet the customer
• How to authenticate (what data to ask for)
• What to say for each status: in-transit, out-for-delivery, delivered, exception, returned
• When to escalate to a human
• How to end the conversation

Exercise: Write your SOP in a document. Compare it with the WISMO agent page when done — how does yours differ?

Step 2: Create the workflow (20 min)

In the AgentFleet Dashboard:

1. Create a new workflow → name it “QuickShip WISMO”
2. Add a Start node → this receives the inbound query
3. Add an Agent node → select the default_sub_agent template
   • Set the model to gemini-2.5-flash
   • Paste your SOP as the system prompt
4. Add a Tool node → assign the fetch_information tool (fetches order data from TMS)
5. Add an End node
6. Connect the nodes → Start → Agent → End, with the Tool node available to the Agent

Step 3: Wire up the tool (15 min)

The fetch_information tool calls the ProjectX API to get consignment details. In the demo environment:

1. The tool is pre-configured — you just need to assign it to your agent node
2. Test it: trigger the workflow with a sample input:

   {
     "input_params": {
       "query": "Where is my order QS-2026-001?"
     }
   }
3. Check the job result in the Dashboard — did the agent call the tool? Did it get data back?

Step 4: Run a basic eval (15 min)

Test your agent with these 5 scenarios:

For each scenario, record: Did it work? What went wrong? What would you change in the SOP?

Step 5: Iterate (15 min)

Based on your eval results:

• Refine the system prompt (SOP)
• Adjust tool assignment if needed
• Re-run failing scenarios
• When all 5 pass, you’ve got a working WISMO agent

Checklist

• SOP document written
• Workflow created in Dashboard with 4 nodes
• Tool assigned and tested
• 5 eval scenarios run
• At least 4/5 scenarios pass

What you learned

• How agents are workflows, not code
• How SOPs translate to system prompts
• How tools give agents access to real data
• How evals tell you whether your agent works

Next steps

• Lab 2: Write a discovery questionnaire
• WISMO agent deep dive
• Eval framework for building comprehensive eval sets

Changelog

• 26 May 2026: Full lab content with step-by-step instructions and eval scenarios.