Your First Agent

Now you will build a small but real agent: it reads a task, classifies urgency, and outputs an action plan. This example is intentionally narrow so you can ship it today.

Project setup

Create a minimal TypeScript app with one model client and one tool.

pnpm add openai zod

Use server-side environment variables only. Never expose keys in client bundles.

import 'server-only';
import OpenAI from 'openai';

export const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY! });

Define a strict output schema

Your agent should return structured JSON, not free text.

import { z } from 'zod';

export const triageSchema = z.object({
  priority: z.enum(['low', 'medium', 'high']),
  owner: z.string().min(1),
  nextAction: z.string().min(10),
  dueInHours: z.number().int().min(1).max(168),
});

export type TriageResult = z.infer<typeof triageSchema>;

Add one tool

Keep tool scope small: normalize due dates and estimate urgency from tags.

type TaskInput = { title: string; tags: string[]; customerTier?: 'standard' | 'vip' };

function scoreUrgency(task: TaskInput): number {
  let score = 0;
  if (task.tags.includes('outage')) score += 5;
  if (task.tags.includes('billing')) score += 2;
  if (task.customerTier === 'vip') score += 3;
  return score;
}

Run the agent

Prompt for JSON matching triageSchema, parse it, and reject invalid output.

const parsed = triageSchema.safeParse(modelOutput);
if (!parsed.success) {
  throw new Error('Invalid agent output');
}
return parsed.data;

Practical advice: log raw model output before parsing, but never log secrets or customer PII.

What to test before shipping

• Empty task title
• Unknown tags
• Very long user input
• Model returns malformed JSON
• Tool returns unexpected values

If these cases fail gracefully, you already have a production-ready foundation for simple internal automations.

Exercise

Build this triage agent and run it against ten real tasks from your backlog. Compare agent recommendations with your manual decisions. Measure agreement rate and edit your scoring logic before adding more tools.