HubSpot has been shipping agents for two years: Prospecting Agent, Customer Agent, Deal Loss Agent, each living in its own corner of the platform. The pattern across portal engagements has been consistent. Teams activate one agent, forget the others exist, and nobody has a single view of what the AI in their portal is actually doing.
Agent Hub is HubSpot's answer to that sprawl. It is not a new agent. It is the management and construction layer for all of them, and the construction part is the bigger story.
This is the first of two Field Notes entries on the release. This one covers what Agent Hub is and how its pieces fit together. The second covers what has to be true about your portal before you should publish anything in it.
Agent Hub is a single console where HubSpot's AI agents are viewed, activated, and managed, organized by the outcomes they drive rather than by hub or product line. Featured agents appear with live status and recent results, grouped under go-to-market goals: building demand, winning deals, delighting customers, scaling growth.
The public beta opened on July 23, 2026 and is included for all Professional and Enterprise portals across Marketing, Sales, Service, Data, and Content Hubs, plus Smart CRM at those tiers. No separate pricing tier was announced at launch. You reach it under Agents in the main navigation.
The consolidation matters because agents that operate in isolation create the failure mode HubSpot itself describes: a prospecting agent emailing an account the same week a service agent is handling that account's open complaint, with neither aware of the other. One console, shared customer context.
Custom agents can "analyze data, generate outputs, and take actions based on your business processes."HubSpot Knowledge Base ยท Create and customize agents in the agent builder July 2026
Agent Builder is a no-code canvas inside Agent Hub where you create custom agents by describing what they should do in natural language, then configuring what they can access and act on. A custom agent has four components, and understanding them is most of understanding the product.
Instructions define the agent's role, goal, approach, and expected output. This is the prompt layer, and specificity here determines output quality everywhere else.
Actions define what the agent can do: read and write CRM records, browse the web, generate content, and connect to external systems through MCP, the Model Context Protocol.
Knowledge defines the standing context the agent draws from on every run: Brand Kits, Ideal Customer Profiles, Products and Services, uploaded documents, and knowledge vaults.
Inputs define the runtime data each individual run acts on, such as a specific deal record or a user-provided value. Knowledge is permanent context; inputs vary per run.
Agent Hub includes its own workflow builder that puts agents and workflows on the same canvas for the first time, with triggers that extend well beyond schedules and enrollment lists. Workflows can fire on contact property updates, webhooks, Slack events, and third-party application events, and can execute actions in external tools like Google Sheets, Asana, and Jira alongside standard HubSpot actions.
Agents run as steps inside these workflows. You can drop an existing agent into a flow or create a new one from a prompt without leaving the canvas, and Breeze Assistant can generate custom code actions on request.
Two mechanics differ from the classic workflow editor. Each workflow takes exactly one trigger, and enrolled records do not re-enroll by default. Both constrain design, and both function as safety rails against runaway enrollment loops.
The structural shift is the point: agents and workflows are no longer separate systems. Deterministic automation and agentic reasoning now compose on one surface.
Custom agents consume HubSpot Credits each time they perform configured actions, drawing on the credits already included in Professional and Enterprise subscriptions. Testing an agent in the builder is free and consumes no credits. Estimated credit costs per run appear in the agent's run history, and monthly run limits can cap consumption per agent.
The practical read: the capability is already paid for, but usage is metered. An agent's trigger scope is now a billing decision as much as an automation decision, which is why the run limit exists and why it belongs in your launch configuration rather than your post-invoice cleanup.
Creating agents requires Super Admin access or the Breeze Studio permission, after a Super Admin enables generative AI access and data sharing in AI Settings. Specifically, toggles for CRM data, customer conversion data, and files data must be on before the builder functions.
Agent access is then assignable per agent: owner-and-admin only, everyone-can-run, or custom team and user scoping. Feature permissions still apply on top. An agent that publishes landing pages can only be run by users who hold landing page publish permissions themselves.
Agents also ship with guardrails that require human approval on actions until you deliberately release them. The approval queue doubles as a record of how the agent interprets your instructions, which makes it valuable well past the first week.
What this supersedes.
HubSpot's agents live in separate corners of the platform with no shared view. Automation means workflows a human designed step by step. AI capability is evaluated agent by agent, activation by activation.
Every agent sits in one console with live status and shared customer context. Agents and workflows compose on a single canvas with event-driven triggers. AI capability is a platform layer, metered in credits and governed by permissions.
The first session is mechanical. Confirm your tier (Professional or Enterprise on any hub), have a Super Admin enable the AI Settings toggles, and confirm who holds Breeze Studio permission. Open Agent Hub under Agents in navigation, review which featured agents are already available to your portal, then open Agent Builder and create one deliberately small custom agent: a single record type, read-only actions, a narrow instruction set.
Test it. Testing is free, so iterate on instructions until the output is consistent. Review the estimated credits in run history, set a monthly run limit, and leave human approval on.
Then stop before publishing anything with write access, because Agent Builder makes one large assumption about your portal: that the data underneath it is in order. Whether that assumption holds for your portal, and what to check before you find out the expensive way, is the subject of the next Field Notes entry.