Notion Custom Agents Go Live: A Million Agents Automating Team Workflows

Notion launched Custom Agents in beta in February, and within two months users created more than one million agents. What surprised the team most was how quickly teams moved from experimenting with a first agent to automating end-to-end workflows that whole companies run on. On May 4, Custom Agents left beta and moved to usage-based pricing. Here is what Notion learned building alongside its users.

The fastest teams treated Custom Agents like internal products

Building an agent is not a one-and-done task. The teams that scaled fastest started with a small group of people who knew their workflows well, iterated until each agent met its success criteria, and shared what they learned. Every agent made the next one easier, and that knowledge compounds quickly when people build together.

At Ramp, that approach grew into more than 300 Custom Agents. Ben Levick, their Head of Ops and Internal AI, describes the momentum: there are two "dopamine hits" — one when you crack the problem and automate work you don't like doing, and another when, on the other side, things get a little quieter and you have more time in your day.

Agent-building became a new operational skill

The teams that built the most durable agents were the ones closest to the problem every day and who wanted a culture of no busywork — not necessarily the ones with the most technical resources. None of these agents were built by engineers:

  • Heidi's product team saves up to 60 hours each month with an internal Q&A agent.
  • Equals Money's product marketing team built a Knowledge Base Amendment agent so docs stay current after every release.
  • Notion's own Customer Support team built an agent that helps reps respond faster and spend more time on complex questions.

If you can describe what you would tell a new hire to do, you can build a Custom Agent.

The biggest wins came from rewiring workflows

Many teams went further than task automation by redesigning how work flows end to end. At Braintrust, competitive intelligence used to require manual research every morning. The marketing team built a Competitive Intelligence Buddy that scans competitor changelogs and Gong transcripts each morning, updates the sales one-pager in Notion, and flags launches that matter for active deals. Now reps show up ready without extra prep.

Cost efficiency comes from both small and big wins

Making the same workflows cheaper over time is one of Notion's core commitments. In the past two months, Custom Agents became up to 50% cheaper to run. The biggest driver was introducing lighter-weight models like Haiku 4.5 and GPT-5.4 Mini, which use up to 10x fewer credits. Because Notion isn't locked into a single model, you can match the right model to the task so you never pay for more than the work needs.

Scaling confidently is the hard part

Teams that scaled agents wanted control at scale without adding a new layer of management. Notion now offers per-agent credit limits, full visibility into every agent on a dashboard, proactive alerts as you approach your credit limit, and the ability to pause agents.

Usage-based pricing

Starting May 4, Custom Agents run on Notion credits as an add-on to Business and Enterprise plans. Recurring monthly credits are $10 per 1,000 credits and flex to your needs, so you can scale up or down monthly.

Getting started: five steps to your first agent

  • Step one, pick the most repetitive task. Don't start with "automate everything." Choose one thing that takes you two or more hours a week with fixed steps — tidying up OKR progress, logging customer requests from email into a database, or aggregating weekly reports scattered across pages.
  • Step two, write the steps out in plain language. If you can clearly write "new request arrives → assess priority → update the project database → notify the owner," you already have what it takes to build an agent.
  • Step three, clean up your data first. An agent's quality ceiling is set by your data structure: databases with consistent fields and naming let agents read and write reliably; messy pages will derail even a smart agent.
  • Step four, get it working on a lightweight model. Validate the flow with a credit-cheap model (like Haiku 4.5 or GPT-5.4 Mini) first, then upgrade only if the task needs it — never pay for capability you don't use.
  • Step five, set boundaries and watch. Give the agent a credit limit, keep the usage dashboard open, and review output quality daily for the first two weeks, feeding failures back into the instructions. That beats trying to get it perfect in one shot.

An example: a weekly-report agent

Say you compile team weekly reports every Friday. Create a "Weekly Report" database with fields for member, completed this week, plans for next week, and blockers. Then have the agent: ① scan the project database and task statuses on Friday afternoon; ② turn each member's changes into a short paragraph; ③ write it into the report database and @mention members to confirm. That used to be hand-copied by the lead; now the agent finishes in ten minutes, and every field points back to the source database, so you can trace the origin with one click.

FAQ

  • Will the agent mess up my data? Start with read-only permissions plus manual confirmation, then open up write access once it's stable. Notion keeps version history, so mistakes can be rolled back.
  • Will credits run out quickly? They can. Set a per-agent credit limit first, watch a week of usage on the dashboard, then decide between a lighter model and lower trigger frequency.
  • How do agents divide work with Zapier / Make? Plain "when A happens, do B" automation is cheaper and more predictable for simple triggers; tasks that require reading, understanding, and judgment are what agents are for. The two complement each other.

Reference: Notion Custom Agents help docs https://www.notion.com/help/guides/custom-agents

16IDC perspective

For companies that rely on Notion for project management, knowledge, and collaboration, Custom Agents mean productivity tools are moving from passively recording work to actively executing it. To use them well, start with a solid Notion project management foundation: clean workflow databases are a prerequisite for stable automation. Remote and hybrid teams can review remote team tool comparisons to decide where agents fit versus other automation. See more in the Business Tools category.

Source: https://www.notion.com/blog/what-we-learned-during-the-custom-agents-beta