How to build your first AI agent, from first goal to going live

HubSpot has published a step-by-step guide to building a first AI agent, and its most useful advice comes before any building starts: check whether the job needs an agent at all. A 2026 Menlo Ventures survey of 5,000 US adults found that one in four AI users has used an agent, so more people are trying them.

What an agent is

An AI agent is software that uses an AI model to work toward a goal without a person guiding each step. It can pick among approved tools, look at the results, and decide what to do next until it finishes or hits a set limit. A marketing example: review a campaign's results, find likely causes for a change, and recommend a fix.

What a first agent needs

The guide credits Elle Oshiro, a professor at HubSpot Academy, with four building blocks:

• A repeatable job the team can describe clearly
• Relevant data that exists, is current, and the agent can access
• A clear starting point, such as a manual request, a schedule, or a change to a CRM record
• A measurable result that defines what a correct answer looks like

For the result, the guide suggests a report that flags weak channels and recommends one change to test. Each recommendation should cite the metrics behind it and admit where data is missing.

When not to use an agent

Agents suit tasks that repeat, are tedious, and require interpreting information and deciding what to do next. If you only need a summary or some analysis of something you paste into a chat, a single AI prompt may be enough. The guide notes that Anthropic recommends starting with the simplest approach that meets the need, because agents can add cost and delay.

If the next steps are always clear, rules-based automation fits better. The guide says even a complex lead-routing process can work this way. A workflow can check territory, company size, and product interest, assign the lead to an eligible rep, and alert a manager if nobody accepts it in time.

How to start

The first step is to write down the manual task you want to hand off, such as reviewing lost deals and recommending content that answers buyers' open questions. Then put together a short brief covering the agent's goal, scope, and expected output. After that, write clear instructions and test the agent before full deployment.

HubSpot users can use a no-code Agent Builder, where you describe the task in plain language. The guide also mentions pre-built agents for jobs like data management and prospecting. The provided text cuts off before the detailed steps on testing, improving, and setting guardrails, so those specifics are not covered here.