Notion Custom Agents: What They Can (and Can't) Do

Notion quietly turned its workspace into an agent platform this year, and on September 9, 2026 it shipped a small but telling update: workspace owners can now control which AI models Notion Agent and Custom Agents use, and set a default model for Custom Agents. If you live inside Notion and have been eyeing Custom Agents as a lighter alternative to a Claude Code subagent workflow, this is a good moment to look at what they actually do, and where they stop.

What Notion Custom Agents actually are

Notion introduced Custom Agents in February 2026, describing them as "autonomous AI teammates that handle recurring work 24/7 across Notion, Slack, Mail, Calendar, Figma, Linear, and your custom MCP servers."

The setup flow is deliberately hands-off. You describe what you need in plain language, set a trigger or a schedule, and the agent writes its own instructions and configures which tools it uses. There is no prompt engineering step for the end user, at least on the surface.

Notion names three primary use cases for Custom Agents:

• Q&A agents that answer repeated questions automatically using your existing knowledge base
• Task routing agents that capture incoming requests, enrich the data, and route them to the right team member
• Status update agents that compile reports pulling from Notion, your connected tools, and the web

Adoption numbers from the launch post give a sense of how far this has already gone: Notion says it now runs more agents internally than it has employees, early testers created more than 21,000 agents, Ramp runs 300+ of them, and Remote reported saving 20 hours a week on task routing alone.

Access, spending, and admin controls

Custom Agents launched as a free public beta through May 3, 2026. From May 4, they started consuming Notion Credits, a usage-based add-on, and are only available on Business and Enterprise plans (existing AI features like Notion Agent, AI Meeting Notes, and Enterprise Search stayed included at no extra cost).

Because agents that act 24/7 across your tools are also agents that can rack up cost and risk quietly, Notion shipped admin guardrails alongside the feature. A May 5, 2026 release added:

• Access control so admins decide who can create Custom Agents, by individual or user group
• Per-agent credit limits set jointly with the agent's creator, plus a workspace-level cap for Enterprise
• A credits dashboard showing usage by agent, spending trends, status, and recent activity
• Email and in-app alerts as the workspace nears its credit limit, with automatic pausing if credits run out or if a new agent starts spending unusually fast
• The ability for admins to disable an agent from Settings, Notion AI, Agents, and work with its creator before re-enabling it

The original launch post also lists reversible changes, logged runs showing every trigger and action, granular permissions, and prompt injection detection guardrails as part of the safety story.

The September 9 update: choosing your own model

The newest change is narrower but matters for anyone running agents at scale. Workspace owners on Business and Enterprise plans can now choose which AI models are available separately for Notion Agent and for Custom Agents, and set a default model for Custom Agents. The setting lives in Settings, Notion AI, General, Model controls.

Notion frames the benefit plainly: it helps "keep costs predictable and support your compliance needs while your teams continue building." Notion has not published which individual models are available in this control, so treat model choice as workspace-configurable rather than something you should assume includes any specific provider.

Where this differs from a Claude Code subagent workflow

If you already run subagents in Claude Code, the contrast is useful. A Claude Code subagent is a file you write and check into your codebase, an ".md" config with its own system prompt and an explicit, scoped set of tools you assign to it. You decide exactly what it can touch, and the main Claude Code session delegates a specific task to it inside a working session, then reads back what it did.

Notion Custom Agents flip that division of labor. You describe an outcome, and the agent writes its own instructions and picks its own tools from what is available (Notion, Slack, Mail, Calendar, Figma, Linear, your MCP servers). That is faster to set up for a non-technical teammate, but it also means the agent's actual behavior lives inside Notion's product rather than in a file you can read, diff, and version.

The two are not really competing for the same job. Custom Agents are built for recurring, scheduled work sitting on top of your existing Notion knowledge base, answering questions, routing tickets, compiling status reports. A Claude Code subagent workflow is built for building things: reviewing code, running a multi-step build, shipping a feature. If your work is "keep this workspace running smoothly," Custom Agents are the more native fit. If your work is "build and ship software," a structured subagent workflow gives you the control and auditability Notion's agents were not designed for.

Notion custom agents for Product Managers

For a PM who already lives in Notion for specs, roadmaps, and meeting notes, Custom Agents are worth a real look before reaching for anything heavier. The status update use case is the most obviously useful one: an agent that compiles a weekly roadmap digest from your Notion pages and connected tools removes a recurring chore with almost no setup. Task routing agents can also take the manual triage out of an intake database, feature request board, or bug tracker, enriching each item and assigning it before you even open it.

Where it gets more interesting is the boundary. Custom Agents are excellent at working with information that already lives in your workspace. They are not built for the kind of work a PM increasingly needs from AI in 2026: turning a rough idea into a working prototype, reviewing a PRD from multiple stakeholder angles with a repeatable process, or synthesizing raw discovery calls into a structured insight. That is a different category of work, closer to building than to reporting, and it is exactly what a Claude Code subagent workflow is designed for.

The takeaway

Notion Custom Agents are a real, maturing product, not a beta gimmick. In seven months they went from free beta to a credit-metered feature with admin spending controls, access restrictions, and now workspace-level model selection. For PMs, they are a strong default for keeping a Notion workspace itself running on autopilot. For actually building product, a subagent workflow with explicit, versioned instructions still does a different, complementary job.