HubSpot Agent Builder: When One Sentence Builds a Marketing Agent – and Who Approves It

Not long ago, a software robot for marketing needed a developer, a service provider or at least patience with a workflow editor. Now you describe in one sentence what the agent should do and get a configuration proposal back. HubSpot offers exactly that: the "Breeze Assistant" builds a custom AI agent in the "agent builder" from a description in plain language.

For mid-sized companies that already use HubSpot this sounds like a shortcut. It is also the moment that decides whether an agent with limits is created, or an agent somebody built "quickly" that nobody oversees afterwards. This article goes through what the official HubSpot documentation actually describes, what it leaves open and which approval rules you should set before the first agent goes active.

The source is the HubSpot knowledge base page "Use Breeze Assistant to build custom agents and agentic automation in the agent builder", last updated October 21, 2026. Where the page says nothing, we say so. We have assigned the topic to our analytics agent Daniel and, where outreach is concerned, to our outreach lead Kai.

What the HubSpot documentation describes

According to the documentation, the path to an agent has five steps. You open the Breeze Assistant in the top navigation bar, enter a description and send it. The assistant proposes a configuration made up of a goal, tools and knowledge sources that you can review and refine. It may ask clarifying questions. When you confirm, a link to the agent in the agent builder is created, where you customize it further.

In the agent builder you can adjust instructions, add actions and test the agent before publishing it. The output of an agent is then reviewed in the "agent inbox", the page says. That is the most important point for our topic.

The feature is available, according to the page, in Marketing Hub, Sales Hub, Service Hub, Data Hub and Content Hub on Starter, Professional or Enterprise, and in Smart CRM and Revenue Hub on Professional or Enterprise.

Creating and customizing agents requires Super Admin or agent builder permissions. Some goals need further permissions; the page gives Edit and Publish permission for generating landing pages as an example. A Super Admin must also enable several AI settings: access to generative AI tools and features, CRM data, customer conversion data, files data and Breeze Assistant access.

That is a sound permission logic, and it transfers to your own approval process: not everyone may build agents, and what an agent may touch depends on what has been enabled in the settings.

What the documentation does not say

• Beta or generally available: The page gives no status. We cannot say whether the feature is considered stable.
• Region: The page names no regions. Whether everything is available in Germany, Austria and Switzerland to the same extent is open.
• Approval before activation: The page says you can test the agent before publishing. Whether an agent can go active without an additional manual review is not stated.
• Price: Apart from the plans that include the feature, no cost is named.
• Limits: No information on the number of agents, usage limits or which AI models sit behind it.

We therefore treat the topic as something that is coming toward companies and should be checked, not as a proven standard function.

"Easy to build" says nothing about whether an agent is set up correctly. The easier an agent is to create, the more important the rule about who may activate it.

The real risk: agents nobody consciously decided on

That an agent can be created by voice or text is a relief and a risk at once. The relief: departments no longer wait for IT. The risk: agents appear for which nobody asked what happens if they are wrong.

A thought experiment: a marketing lead at a machinery manufacturer with 120 employees asks the assistant to build an agent that follows up with new prospects from a trade-fair contact import. The proposal looks plausible, she confirms, tests with two contacts and switches it on. Two weeks later it turns out that existing customers with open complaints also received a promotional message because the target group was too broad. Every single step was reasonable. What was missing was a second person who read the target group and the action before activation.

The agent inbox mentioned in the documentation is a good starting point: results are reviewed there. Whether that means a review before or after execution is not clear from the page, and that is the question to clarify in your own instance.

Three approval rules to set before the first agent

1. Who may create agents, who may activate them?

The documentation distinguishes Super Admin rights from the "agent builder" permission. Use that to separate creating from activating: one person builds, another approves. It costs minutes and prevents the classic case of the author approving their own work.

2. Which action may an agent perform without asking?

Distinguish reversible from irreversible actions. A draft can be discarded; a message sent to 3,000 contacts cannot. Decide that everything visible externally is seen by a human before sending. That is the principle we apply ourselves: emails are created as drafts and not sent directly.

3. How do you measure whether the agent works well?

A test with two contacts is a functional test, not proof of quality. Decide in advance what "good" means. A simple rule of thumb: check every result in the first weeks, then samples, and extend freedom only step by step.

Three example tasks and what to watch

• Internal summary: An agent summarizes weekly which inquiries arrived. Damage in case of error is low because nothing goes outside. This is the obvious first use case.
• Landing page draft: The example comes from the documentation and requires additional edit and publish rights. We would let the agent create drafts only and reserve publishing for a person who reads text, promises and legal aspects.
• Message to contacts: Anything that reaches people outside your company belongs in the highest review level. Target group, text and timing should be seen by a human before sending.

These are examples for classification, not statements about what the agent builder can do in detail.

Assignment: Daniel and Kai

On our website we describe Daniel as the analytics agent: "Numbers without commentary are useless. Daniel makes data decidable." That fits the question of how to evaluate an agent's work. An agent that contacts people or creates landing pages generates data, and without evaluation nobody knows whether it works well. We introduce Kai as outreach lead, responsible for strategy, sequences and pipeline. The assignment is our own classification; HubSpot does not describe the feature this way.

How we proceed in the audit

1. Collect use cases: Which weekly marketing and sales tasks are clearly defined enough for an agent?
2. Clarify data access: The documentation names several AI settings a Super Admin must enable. We check with you which ones the agent really needs and leave the rest off.
3. Classify actions: Reversible or not, internal or externally visible?
4. Define the review routine: Who sees the results, how often, and what triggers a correction?
5. Plan expansion stages: The agent gets more freedom only once the review routine has confirmed it over an agreed period.

What you can prepare now

1. Check your instance: See whether the Breeze Assistant and agent builder appear and which plans apply.
2. Review rights: Clarify who is Super Admin and who has the agent builder permission.
3. Document settings: Note which AI settings are active today and why.
4. Pick a first use case: Choose a task with low damage in case of error.
5. Write down the review rule: Who sees results and when is the agent stopped?

An agent with limits and an approval step

An agent without limits often arises from good intentions: someone wants to save time, describes a task, confirms the proposal and switches it on. After that it works until a mistake is noticed. An agent with limits and an approval step has three things: a named person who approves it, a list of what it may do without asking, and a review routine that watches its results.

For us this is everyday practice: emails are created as drafts and not sent directly, and content does not go live without a brand check by Magdalena. The same logic applies to any agent, whatever platform it is created in. The HubSpot function makes building easier. The responsibility for giving it limits stays with you.

Source: HubSpot Knowledge Base, "Use Breeze Assistant to build custom agents and agentic automation in the agent builder", last updated October 21, 2026 (primary source).