Build Your First AI Agent: Idea to Live in a Weekend

You don't need a machine-learning degree or a funding round to ship a working AI agent. You need one annoying, repetitive task and a free weekend. This guide takes you from a blank idea to a live agent answering real messages, using a single concrete example you can copy.

What an "agent" actually is

Strip away the hype and an AI agent is three things wired together:

• An LLM (the brain — Claude, GPT, Gemini) that reads input and decides what to say or do.
• A few tools (the hands — look up an order, send a WhatsApp reply, save a row to a sheet).
• A trigger (the doorbell — a new WhatsApp message, an incoming email, a scheduled time).

That's it. A chatbot just talks. An agent talks and does — it can call a tool, read the result, and act. Your first agent should stay close to "smart auto-responder" and grow from there. Don't build Jarvis. Build the thing that answers the same five questions your customers ask every single day.

Pick a tiny first project that solves one real pain

The most common mistake is picking something too big. The right first project fits all four of these:

• It solves one real, recurring pain (not five "nice to haves").
• A human currently does it in under 2 minutes per instance.
• It has clear right answers (so you can tell when the agent is wrong).
• Getting it slightly wrong is cheap (no money moves, no legal promises).

Good first agents for an Indian small business:

We'll build the WhatsApp FAQ bot — it's the highest-leverage starter for most Indian SMBs, where customers message in Hinglish at all hours.

The 5 building blocks

Every agent is the same five blocks. Design these on one page before you write any code.

1. Trigger — what wakes the agent up. (Incoming WhatsApp message.)
2. Prompt — the instructions + personality + rules. (Your system prompt.)
3. Tools — what it can fetch or do. (Read FAQ doc; hand off to human.)
4. Memory — what it remembers. (This conversation's last few turns; maybe the customer's name.)
5. Output channel — where the answer goes. (Back to the same WhatsApp chat.)

If you can fill in those five lines for your idea, you can build it.

Worked example: a WhatsApp FAQ agent, end to end

The pain: Sharma Sweets gets 40+ WhatsApp messages a day asking the same things — timings, delivery area, today's specials, whether they take UPI. The owner's son answers them between serving customers.

The five blocks, filled in:

Step 1 — Write the knowledge, not the code

Put everything the agent should know in one short doc. This is 80% of the quality.

SHOP: Sharma Sweets, Karol Bagh, Delhi
HOURS: 9 AM - 9:30 PM, all days
DELIVERY: Within 4 km, free above ₹500, else ₹40
PAYMENT: Cash, UPI (sharmasweets@upi), cards
SPECIALS: Update daily — today: Kaju Katli ₹900/kg, Motichoor Laddoo ₹420/kg
GST: Prices include GST; invoice on request
DO NOT: Promise delivery times during festivals; quote bulk/wedding orders — escalate those.

Step 2 — Write the system prompt

This is the agent's brain. Copy, paste, and edit the bracketed bits:

You are the WhatsApp assistant for [Sharma Sweets], a sweet shop in [Karol Bagh, Delhi].

YOUR JOB: Answer customer questions using only the SHOP FACTS below.
Be warm, brief (1-3 sentences), and reply in the same language the
customer uses — including Hinglish. Use ₹ for prices.

SHOP FACTS:
[paste the knowledge doc here]

RULES:
- If the answer isn't in SHOP FACTS, say you'll check and call escalateToOwner().
- Never invent prices, timings, or delivery promises.
- For bulk, wedding, or custom orders, say a team member will confirm
  and call escalateToOwner().
- Don't collect payment or share anything beyond SHOP FACTS.
- If a customer seems upset, apologise once and escalate.

Always end with a small helpful nudge, e.g. "Aur kuch chahiye?"

Two rules do the heavy lifting: only use the facts (stops hallucination) and escalate when unsure (makes mistakes safe).

Step 3 — Wire it up

You have three honest options, fastest first:

• Buy a hosted one. Subscribe to a ready WhatsApp FAQ agent, paste your knowledge doc, connect your number. Live in an hour. Browse AI agents.
• Buy the source code and self-host if you want to own and customise it.
• Build from scratch: WhatsApp Business / Cloud API → a small webhook (Node or Python) → call the LLM with your prompt + last 6 messages → send the reply back. For the deployment mechanics, see Deploy & Host AI Agents the Right Way.

For a first weekend, start hosted. You'll learn what "good" looks like before you sink time into plumbing.

How to test it before anyone relies on it

Never point a fresh agent at live customers. Run it through a script first. Build a list of 20 real messages (pull them from your actual WhatsApp history) covering the boring, the tricky, and the out-of-scope.

• 10 normal questions (timings, delivery, payment, today's price)
• 3 Hinglish / typo messages ("delivry hoti hai kya bhai")
• 3 out-of-scope ("do you have a job opening?") — should escalate, not guess
• 2 "trap" questions with no answer in the doc — must NOT invent
• 1 angry message — must apologise + escalate
• 1 bulk/wedding order — must escalate, not quote

Score each reply: Correct / Wrong / Made-something-up. The "made-something-up" column must be zero before you go live — a confident wrong price is worse than no answer. If it invents things, tighten the only use SHOP FACTS rule and re-test. For sharper prompt fixes, Prompt Engineering for Real Business Tasks and the 40 WhatsApp Customer-Support Prompts (Copy-Paste) are worth a read.

Going live + what to monitor in week one

Go live small. Turn the agent on for a few hours a day, or let it draft replies that you approve before sending, then graduate to fully automatic once you trust it.

Watch these for the first seven days:

Each evening, skim the day's chats for 5 minutes. Every question it fumbled is a one-line addition to your knowledge doc — your agent gets visibly smarter overnight. Keep a human reachable via the escalation path; that safety net is what lets you sleep.

One India note: if your agent stores customer names or numbers, you're handling personal data under the DPDP Act — collect only what you need, tell people a bot is replying, and don't hoard chat logs you don't use. A line like "Hi, you're chatting with our assistant bot" at the start keeps you honest and builds trust.

Next steps

• Write your one-page knowledge doc and the system prompt above — today, before any code.
• Build your 20-message test script from real WhatsApp history and run it.
• Start hosted: browse a ready WhatsApp agent on AgentDukaan and go live this weekend.
• Once it works, consider listing your own version for other shops — Sell Your First AI Agent on AgentDukaan.

When you're ready, AgentDukaan (agentdukaan.in) has hosted agents, source code, and done-for-you setup — pick whichever gets your first agent live fastest, and the Help Center is there if you get stuck. No rush; the best first agent is the one you actually ship.