Skip to main content
This is a TypeScript port of Pydantic AI’s canonical bank support example. It demonstrates the three defining features of Vibes: typed dependency injection via RunContext, dynamic system prompts via the instructions function, and structured output via outputSchema.

​
What you’ll learn

  • Typed dependency injection with Agent<TDeps, TOutput>
  • Dynamic system prompts with the instructions function
  • Tools that access dependencies via tool<TDeps>()
  • Structured output with Zod schemas and outputSchema

​
How Pydantic AI maps to Vibes

Pydantic AIVibes
SupportDependencies dataclasstype Deps = { customerId: number; db: DatabaseConn }
@support_agent.system_prompt decoratorinstructions: async (ctx) => ... in AgentOptions
SupportOutput(BaseModel)z.object({ supportAdvice, blockCard, risk })
Agent(model, output_type=..., deps_type=...)new Agent<Deps, Output>({ outputSchema, instructions })
agent.run_sync(...)await agent.run(...)

​
Prerequisites

  • ANTHROPIC_API_KEY set in your environment
  • Vibes installed (deno add npm:@vibesjs/sdk npm:@ai-sdk/anthropic npm:zod)

​
Complete example

import { Agent, tool } from "npm:@vibesjs/sdk";
import { anthropic } from "npm:@ai-sdk/anthropic";
import { z } from "npm:zod";

// --- Mock database ---
class DatabaseConn {
  async customerName(id: number): Promise<string> {
    const names: Record<number, string> = { 123: "Alice", 456: "Bob" };
    return names[id] ?? "Unknown";
  }

  async customerBalance(id: number): Promise<number> {
    const balances: Record<number, number> = { 123: 1250.50, 456: 89.0 };
    return balances[id] ?? 0;
  }
}

// --- Dependency type ---
type Deps = { customerId: number; db: DatabaseConn };

// --- Output schema ---
const SupportOutput = z.object({
  supportAdvice: z.string().describe("Advice to give the customer"),
  blockCard: z.boolean().describe("Whether to block the customer's card"),
  risk: z.number().int().min(0).max(10).describe("Risk level 0-10"),
});

// --- Agent ---
const supportAgent = new Agent<Deps, z.infer<typeof SupportOutput>>({
  model: anthropic("claude-sonnet-4-6"),
  systemPrompt:
    "You are a support agent in our bank. Give the customer support and " +
    "judge the risk level of their query. Reply using the customer's name.",
  // Dynamic instructions - called before each run to inject customer context
  instructions: async (ctx) => {
    const name = await ctx.deps.db.customerName(ctx.deps.customerId);
    return `The customer's name is ${name}.`;
  },
  tools: [
    tool<Deps>({
      name: "customer_balance",
      description: "Returns the customer's current account balance",
      parameters: z.object({}),
      execute: async (ctx) => {
        const balance = await ctx.deps.db.customerBalance(ctx.deps.customerId);
        return `$${balance.toFixed(2)}`;
      },
    }),
  ],
  outputSchema: SupportOutput,
});

// --- Run ---
const db = new DatabaseConn();

const result = await supportAgent.run("What is my balance?", {
  deps: { customerId: 123, db },
});

console.log(result.output);
// {
//   supportAdvice: "Alice, your current balance is $1,250.50.",
//   blockCard: false,
//   risk: 1,
// }

​
Run it

deno run --allow-net --allow-env bank-support.ts

​
How it works

Dependency injection: Deps is a plain TypeScript type. Vibes passes it through RunContext to instructions, tools, and validators - no global state, no singletons, fully testable. Swap DatabaseConn for a mock in tests by passing different deps at call time. instructions function: Called before each agent run. Returns a string that is appended to the system prompt. Use this for per-request context (customer name, user preferences, session data). Unlike a static systemPrompt, it can read from the database or any async source. tool<Deps>(): The type parameter makes ctx.deps typed as Deps inside execute. Without the type param, ctx.deps is unknown. The tool receives the full RunContext as its first argument. Structured output: outputSchema instructs the model to return JSON matching the Zod schema. result.output is typed as z.infer<typeof SupportOutput> - no parsing or casting needed. Invalid model output raises a typed error.

​
Next steps