Agent Builder: what it is, how it works, and how to build your first agent

How Agent Builder works, from the four build steps to guardrails, channels and Artifacts, plus which agent to build first for your store.

Agent Builder: what it is, how it works, and how to build your first agent

Your team handles the same few jobs all day. A shopper asks where an order is. Another wants to cancel it. A third can't choose between two products and leaves the page.

An agent can take those jobs on. Agent Builder is how you build one, and this guide covers all of it: what it is, why you'd use it, how each step works, where it runs, and which agent to build first.

What Agent Builder is

Agent Builder is Manifest AI's self-serve tool for building agents that automate marketing and support for your store. You describe the job. You choose the tools the agent can use. You set where it hands over to your team. Then you test it and go live.

An agent has one job and sees it through. It finds the right product for a shopper, looks up an order, cancels it and issues the refund. If the job needs another skill, it passes that piece to a sub-agent. If a person should step in, it hands over.

Agents in Agent Builder call tools. They search your catalog, check stock, track a shipment, cancel an order. The reply comes after the work is done, so the shopper gets an answer that is true for their order.

Agent Builder showing an agent that calls tools before it replies

Why use Agent Builder

  1. Your team gets the repeat work back. Order status, cancellations, product picks. The agent takes the first pass at these jobs, so your team spends its time on the conversations that need a person.
  2. You build it yourself. Every step is written in your own words, and a developer isn't part of the process.
  3. You decide the limits. Guardrails, handover rules and a test before go-live come with every agent.
  4. One build runs on four channels: your storefront, Instagram, WhatsApp and email.

Busy seasons repeat the same jobs at higher volume. An agent you built and tested in a quiet week is already working when the rush arrives.

The four steps

Every agent is built in 4 steps. We'll use a gift finder as the example.

1. About this agent

Give the agent a name and one job. "Help shoppers find a gift" works. "Handle customer service" is too wide to test. A narrow job is easier to get right, and you can always build a second agent.

Step 1 in Agent Builder: naming the agent and giving it one job

2. When it steps in

Pick the moment the agent acts. A shopper asks for gift ideas. A shopper asks where an order is. A shopper asks to cancel. The agent acts on the moments you choose and leaves the rest to your team.

Step 2 in Agent Builder: choosing when the agent steps in

3. How it responds

This is where you write the work. Brief the agent the way you'd brief a new hire: ask who the gift is for and the budget, search the catalog, suggest 3 products that are in stock.

Type @ to pull a tool from the library. Search products, Check stock and Pull reviews are there for agents that help before the order. Track shipment is there for agents that help after it. If you use a system we don't list, connect it over a REST API. When part of the job is its own task, hand it to a sub-agent.

Step 3 in Agent Builder: writing how the agent responds and pulling in tools with @

4. Handover rules

Decide which conversations go to your team. Angry customers, high-value orders, refunds above a limit you set. The agent finishes everything else.

Step 4 in Agent Builder: setting handover rules

Test it, then go live

Run the agent on real shopper messages first. Nothing reaches a shopper until you go live, and you can pause the agent at any time.

Testing an agent on real shopper messages in Agent Builder
Going live with an agent in Agent Builder

Tools and sub-agents

Tools are what let an agent do work, so they come first. The library covers the common jobs: searching products, checking stock, pulling reviews, tracking a shipment. Your own systems connect over a REST API.

Sub-agents handle the parts of a job that are a task of their own. A returns agent can send the shipment lookup to a sub-agent and carry on with the answer. You decide which sub-agents an agent can call.

Guardrails

Every agent follows the rules you write. "Never offer a discount that isn't live on the site." "Refunds over $500 go to a person." Use your own numbers and your own policies. You see how the agent applies them in the test, before a shopper does.

Guardrails in Agent Builder

Where it runs

Build the agent once and it runs on your storefront, Instagram, WhatsApp and email. The rules and the handover stay the same on each channel.

Artifacts

Some jobs need more than a message. Picking a size. Choosing which order to return. Pausing a subscription.

With Artifacts, your agent builds an interactive screen for the job, right on your storefront. You describe the flow in a prompt, and the screen comes out in your brand's own look. A shopper can find a product, start a return or skip a subscription box without leaving the page.

An Artifact built by an agent, shown on a storefront

What you can build

Here are six agents, with a way you could set up each one. They are examples to start from.

AgentWhen it steps inTools it callsHands over when
Gift finderA shopper asks for gift ideasSearch products, Check stock, Pull reviewsThe request is outside your catalog
Product recommendationA shopper is comparing productsSearch products, Pull reviewsThe shopper asks about a bulk or custom order
Size and fitA shopper asks which size to pickSearch products, Pull reviewsThe shopper wants advice on a fit issue after buying
Order trackingA shopper asks where their order isTrack shipmentThe shipment is marked late or lost
CancellationA shopper asks to cancelLook up order, Cancel orderThe order has already shipped
Returns and refundsA shopper asks to return an itemLook up order, Refund orderThe refund is above your limit

Your first agent

Pick one job and run it through 3 checks.

  1. It repeats. Your team does it every day.
  2. It has a right answer. You can say what a good reply looks like.
  3. It has a clear edge. You know when a person should take over.

We'd start with order tracking. It repeats all day, the right answer is in the shipment data, and the handover is easy to define. But the best first agent is the one your team is most tired of doing.

Once you have the job, fill in the brief. It is the same four steps, on one page.

Job:
Moment it steps in:
Steps it follows:
Tools it calls:
Hands over when:

Here is one filled in.

Job: Tell shoppers where their order is.
Moment it steps in: A shopper asks about an order.
Steps it follows: Ask for the order number. Look up the shipment. Share the status and the delivery date.
Tools it calls: Track shipment.
Hands over when: The shipment is late or lost.

Questions people ask

Do I need a developer? You write the steps in your own words, and the tool library covers the common jobs. If you'd like a hand with the first build, we'll do it with you on a call.

What if the agent says something wrong? You set the rules it follows, you choose which chats go to your team, and you test it before it goes live. You can pause it at any time.

Which platforms does it work with? Agent Builder has 30+ integrations ready to use, and you can bring your own over a REST API.

Can it process refunds? Yes, within a guardrail you set. Anything outside it goes to a person.

What does it cost? Talk to us and we'll walk you through it.

Get started

Bring the job you'd hand over first, and we'll help you build the agent on a 30-minute call. Bring your own store and your own job.