How Notion Uses Claude Managed Agents to Build Team Agent Workflows

Notion has long been the place where teams organize their knowledge and coordinate work. As AI agents started to mature, Notion’s product team recognized an opportunity that was also a design constraint: most agent tools are built for a single person and a single agent, which doesn’t fit how organizations actually function. The challenge wasn’t just technical—it was about replicating the collaboration primitives that Notion had already figured out for human teams, and extending them to agents.

The pre-AI bottleneck wasn’t a specific operational failure; it was a structural limitation of the emerging agent ecosystem. Running a coding agent to prototype a feature, or a knowledge-work agent to produce a presentation, required dedicated attention and a one-to-one interface. Teams couldn’t delegate work in bulk, couldn’t review outputs collaboratively, and couldn’t build on shared context in the way they did when using Notion for human collaboration.

Notion product manager Eric Liu and his team built an agent orchestration layer in Notion using Claude Managed Agents, Anthropic’s composable API infrastructure for running long-duration hosted agents at scale. The architecture centers on a task board that acts as an orchestrator: a user creates a task, moves it to “ready to start,” and Claude picks it up—pulling context from connected pages, design docs, API references, and product requirements already in Notion. Users can kick off 30 or 40 tasks at once, and the platform routes completed outputs to the right reviewers for approval. The self-improving skills database is one of its more unusual features: after every merged pull request or completed task, Claude identifies lessons from the work and updates the skills database automatically, so agent quality improves over time without manual curation.

The most vivid demonstration of what this made possible came during early internal testing. Liu had 30 prototyping tasks queued up, dragged them all to “start,” stepped away, and returned to find all 30 completed. What had previously required 12 hours of individual prototyping work was done in about 20 minutes. He then tagged a teammate, and they reviewed and refined the outputs together in Notion in real time—turning what is normally a solo AI interaction into a shared creative moment.

Notion’s roadmap now points toward a deeper rethinking of what the product interface is for. The pages and databases that once helped humans do tasks will increasingly serve as the review layer for agent output—surfacing suggested edits, managing version history, and routing approvals. Liu’s read is that the central interface question in knowledge work is shifting: not “how do you help people get things done?” but “how do you help people review what agents have done?” Notion is positioning itself as the place where that review happens, at team scale.