Testing Customer Support Agents

import scenario from "@langwatch/scenario";
import { describe, it, expect } from "vitest";
import { openai } from "@ai-sdk/openai";
 
describe("Customer Support Fraud Investigation", () => {
  it("should investigate fraud and explore customer account", async () => {
    const result = await scenario.run({
      name: "fraud investigation and card security",
      description:
        "Customer discovers unauthorized transactions on their account and is worried about fraud. They need immediate help to secure their account and investigate the suspicious activity.",
      agents: [
        createBankSupportAgent(),
        scenario.userSimulatorAgent({ model: openai("gpt-4o-mini") }),
        scenario.judgeAgent({
          model: openai("gpt-4o"),
          criteria: [
            "Agent takes fraud concerns seriously and responds with urgency",
            "Agent offers concrete security actions like card freezing",
            "Agent provides clear next steps for fraud investigation",
            "Agent maintains professional and reassuring tone",
          ],
        }),
      ],
      script: [
        scenario.user(
          "Hi, I just checked my account and there are transactions I didn't make. I think my card was stolen!"
        ),
        scenario.agent(),
        checkCustomerExplorationCalled, // Custom assertion - define this to check if explore_customer_account tool was called
        scenario.user(
          "There's an $85 charge at Amazon and a $45 charge at some gas station. I definitely didn't make these purchases."
        ),
        scenario.agent(),
        scenario.user(
          "Yes, please help me secure my account right away. I'm worried about more charges."
        ),
        scenario.agent(),
        verifyNoInappropriateTools, // Custom assertion - define this to verify no inappropriate tools were called
        scenario.judge(),
      ],
    });
 
    expect(result.success).toBe(true);
  });
});

describe("Customer Support Complex Issues", () => {
  it("should use knowledge base for complex multi-part issues", async () => {
    const result = await scenario.run({
      name: "complex multi-issue banking problem",
      description:
        "Customer has multiple interconnected banking problems: locked online banking, unexpected fees, and missing direct deposit. They need systematic help and the agent should use knowledge base guidance.",
      agents: [
        createBankSupportAgent(),
        scenario.userSimulatorAgent({ model: openai("gpt-4o-mini") }),
        scenario.judgeAgent({
          model: openai("gpt-4o"),
          criteria: [
            "Agent addresses all parts of the multi-faceted problem",
            "Agent provides systematic approach to resolving issues",
            "Agent shows empathy for customer frustration",
            "Agent offers clear next steps for each problem",
          ],
        }),
      ],
      script: [
        scenario.user(
          "I have multiple problems with my account. My online banking is locked, there's a $35 fee I don't understand, and my paycheck didn't deposit."
        ),
        scenario.agent(),
        checkMessageSuggestionCalled, // Custom assertion - define this to check if get_message_suggestion tool was called
        scenario.user(
          "I've tried resetting my password multiple times and I really need access to pay my bills. This is really stressing me out."
        ),
        scenario.agent(),
        scenario.judge(),
      ],
    });
 
    expect(result.success).toBe(true);
  });
});

describe("Customer Support Escalation", () => {
  it("should escalate when customer demands human support", async () => {
    const result = await scenario.run({
      name: "customer escalation to human agent",
      description:
        "Customer has been dealing with an ongoing issue and is frustrated. They explicitly demand to speak with a human agent or manager. The agent should handle this professionally and escalate appropriately.",
      agents: [
        createBankSupportAgent(),
        scenario.userSimulatorAgent({ model: openai("gpt-4o-mini") }),
        scenario.judgeAgent({
          model: openai("gpt-4o"),
          criteria: [
            "Agent acknowledges customer's frustration empathetically",
            "Agent offers to escalate when requested",
            "Agent provides escalation timeline and process information",
            "Agent maintains professionalism despite customer frustration",
          ],
        }),
      ],
      script: [
        scenario.user(
          "I've been calling about this same issue for two weeks and nobody can fix it. I want to speak to a real person who can actually help me!"
        ),
        scenario.agent(),
        scenario.user(
          "No more troubleshooting! I want a manager or supervisor right now. This is unacceptable service."
        ),
        scenario.agent(),
        checkEscalationCalled, // Custom assertion - define this to check if escalate_to_human tool was called
        scenario.judge(),
      ],
    });
 
    expect(result.success).toBe(true);
  });
});