When multiple Breeze Agents start operating, management needs arise such as ``which Agent is doing what,'' ``I want to set guardrails,'' and ``I want to add and delete new agents.''Breeze Studio is a control panel for managing and customizing all these Agents and Custom Assistants.is. Breeze Marketplace is a place to discover and install professional agents developed by third parties. This chapter explains all the features of Studio, the main agents of Marketplace, and how to design guardrails.
Breeze Studio is a management screen accessed from "Settings → Breeze → Breeze Studio". You can check the status, change settings, and check the performance of all Breeze Agents and Custom Assistants installed in HubSpot in one place.You can see in a list which agents are running, how much credit is being consumed, and how many inquiries have been processed.。
Breeze Marketplace (starting public beta at INBOUND in 2025) is a catalog where you can discover and install more than 20 professional agents developed by HubSpot, as well as agents developed by third-party partners.We have specialized agents that are outside of the core 4 agents (Content/Prospecting/Customer/Knowledge Base)。
From 2025 to 2026, some Marketplace agents will migrate to GPT-5 architecture.ask for evidenceI am doing it. However, the core 4 agents (Content / Prospecting / Customer / Knowledge Base) are currently operating stably on the GPT-4 architecture, and the transition to GPT-5 is planned to proceed in stages.
What are the most important administrative challenges when deploying Breeze Agent across your organization?"Preventing AI from acting in unexpected ways"That's true. Guardrails define the agent's scope of action, and approval flows provide human checkpoints for important actions.
Run Agent is an action (in early 2026, private beta) that can launch any Breeze Agent within a HubSpot workflow.“Start the Closing Agent when the deal stage changes,” “Have the Customer Agent process the ticket when it is created,” “Have the Prospecting Agent begin research when the form is submitted.”CRM event-driven automation will be realized.
| workflow trigger | Agent to start | Actions performed by the Agent | consumption credit |
|---|---|---|---|
| Deal stage changed to "Proposed" | Closing Agent | Analyze transaction details to draft and send personalized additional proposal emails to agents for closing. | 10cr/time |
| New contact became MQL | Prospecting Agent | Start researching contacts. Start monitoring buyer signals and automatically reach out when signals are detected | 100cr/contact/month |
| A ticket has been created (Category: General Inquiries) | Customer Agent | Search KB and generate draft answers. Automatically send if you are not a VIP customer, suggest to a representative if you are a VIP | 100cr/conversation |
| A new company record has been created | Data Agent | Run Smart Properties. Auto-complete and fill in fields with competitive tools, recruitment status, and ICP scores | 10cr/prompt |
| The transaction became Closed Lost. | Deal Loss Agent | Analyzes the reasons for lost orders and generates improvement proposals for similar projects in the future. Automatically added to weekly Deal Loss report | 10cr/time |
As of March 2026, Run Agent is in private beta. Applications can be made from HubSpot's product roadmap page.Before incorporating it into the production workflow, test it thoroughly in a sandbox environment first.We strongly recommend that you do so. In particular, workflows that involve "automatic sending" run the risk of sending emails to the wrong contacts at the wrong time.
Get an overview of how many Agents are running, how much credits they're consuming, and how well they're performing in Studio. To prevent the situation where you realize your credits have run out, you need to get into the habit of checking Studio's Analytics tab on a weekly basis.
Based on the core 4 agents (Content/Prospecting/Customer/Knowledge Base), we will gradually add specialized agents such as Deal Loss, Customer Health, and Personalization. The correct approach is to ``check the effects one by one and add them'' instead of ``adding everything at once.''
Guardrails prevent the risk of "leaving it to AI and damaging your brand." Be sure to set the following four points: action authority, target contact restrictions, prohibited topics, and approval flow before rolling it out company-wide. Deployment without configuration can become a management risk.
By activating the agent within the workflow, complete automation such as "stage change → closing AI runs" and "lost order → automatic analysis" can be achieved. However, since it is a private beta, we will carefully launch it into production after testing.