Summary: Notion 3.3 shipped Custom Agents on February 24, 2026, letting users assign responsibilities with triggers, schedules, MCP connections, per-agent model choice, scoped permissions, logs, and reversible changes. Custom Agents are free through May 3, 2026, after which runs consume Notion credits. The release turns scheduled agents into governed team infrastructure and shifts AI cost from subscription decisions to measurable team usage.
Notion shipped version 3.3 on February 24, 2026, and one feature carried the entire release: Custom Agents. This isn’t another AI widget parked in a chat box, waiting to be prompted. It’s an agent you hand a job to, bolt a trigger or a schedule onto, and leave running around the clock. Notion’s own framing is blunt: pass the repetitive knowledge work to “AI teammates” and keep human hours for the calls that need judgment.
For developers and product teams, the feature list is only half the story — the business-model turn is the other half. AI capability is stepping out of the bundled subscription and into usage-based billing. From May 4 onward, every agent run burns Notion credits, which means the logic of the API metered bill has arrived, at scale, inside an end-user office suite.
The Job Custom Agents Do
The operating verb is “assign,” not “chat.” Define a responsibility, a trigger, or a schedule, and the agent takes it from there. Notion’s official use-case list includes:
- Q&A agents that field a team’s repeated questions by drawing on Notion, Slack, Notion Mail, Notion Calendar, and MCP sources
- Task-routing agents that catch incoming requests, turn them into tasks, and hand each one an owner
- Status agents that draft standup notes, sprint recaps, and OKR reports on their own
- Email triage and drafting, scheduling across participants, and FigJam-to-Notion page conversion
Every agent also picks the model running underneath it. Model choice first appeared in Notion 3.2; 3.3 pushes it down to the individual agent, which quietly recasts model routing as an everyday setting that non-engineers handle.
Triggers, Schedules, and How Far MCP Reaches
Integrations draw the line on what an agent can touch. Alongside the first-party connectors for Slack, Notion Mail, and Notion Calendar, 3.3 taps Linear, Figma, and HubSpot through MCP, and ships agent templates for common workflows.
Why this design matters: agents don’t sit in a queue waiting for Notion to build connectors one at a time. Point one at an MCP server and it reaches straight into the toolchain the team already runs. For teams with internal MCP infrastructure in place, that’s a schedulable execution surface acquired for essentially zero extra plumbing.
The Admin Controls: Permissions and Audit Trails
What stalls enterprise adoption is rarely agent intelligence — it’s agents taking liberties. Notion’s counter is a concrete set of controls: agents run under a teammate-style permission model with access scoped item by item, can be switched off at any moment, log every single run, and keep changes reversible. Admins on Business and Enterprise plans additionally decide who gets to create agents in the first place.
“Every run leaves a log entry” and “every change can be undone” are the floor-level requirements before an agent graduates from personal toy to team infrastructure. Strip either one away, and no volume of automation scenarios clears a security review.
Free Until May 4, Metered After
The most practical detail in the release is the billing timeline. Custom Agents cost nothing through May 3, 2026. On May 4 the meter switches on: runs consume Notion credits, sold as an add-on to Business and Enterprise plans, with admins buying exactly as much capacity as they need.
That leaves a free runway of roughly ten weeks between now and early May. The sensible play is to spend it standing up the workflows that genuinely pay off and capturing the run data — not to defer budgeting until metering starts, at which point an exhausted credit balance simply switches the agents off.
Numbers from the Beta
Notion also published its beta figures: early testers created 21,000 agents between them, and Notion itself runs 2,800 internally. Among the named cases, Remote puts its savings at 20 hours a week, with triage accuracy above 95% and more than a quarter of tickets resolved without a human in the loop; Braintrust clips 20 minutes off each daily update; Notion’s own security team banks 6+ hours a week with an agent called Scruff. One tester stood up 15 agents in the space of a week.
These numbers come straight from the vendor, so apply the usual discount. Even so, build speed is a signal in its own right: once the barrier drops low enough for 15 agents in a week, the bottleneck moves fast from “can we build this” to “how do we govern it.”
Reading the Release: Notes for Product Teams
Three things stand out. First, productivity suites are converging on the same answer — agents plus MCP — and Notion’s seat is unusually good: it is the knowledge base, the task system, and the document front door all at once, so an agent living there has far more to do than one confined to a chat-only tool. Second, credit billing relocates the AI cost decision, from a subscription call owned by procurement to a usage call made team by team, and the teams that can put a number on value are the ones holding leverage at renewal. Third, governance should run ahead of scale: use the free window to settle naming conventions, permission scopes, and log-review routines, because doing it now costs far less than retrofitting it after May.