Anthropic explains how Notion runs Claude tasks from shared work boards

A shared Notion task board can now serve as the starting point for Claude to write code or produce client deliverables. In an October 12 case study, Anthropic describes Notion’s integration of Claude Managed Agents for long-running work. Notion says teams can delegate complex assignments inside the workspace and route the results to colleagues for review.

Customers connected Claude themselves

The push for a native integration began with customers doing the connecting themselves. After a demo of agents working on a kanban board—a board that organizes tasks by status—went viral, users assembled their own integrations. They used webhooks and custom servers to get Claude working inside Notion, according to the case study.

Eric Liu, a Notion product manager, said that effort showed customers wanted agents within their existing workspace, with Claude their preferred choice. The problem Notion wanted to address was collaboration: a person working privately with an agent gives colleagues little visibility into the output or opportunity to join the work. Notion’s response was to bring agent coordination into the shared workspace.

A board status starts the work

In the workflow described by Anthropic, a team creates a task and moves it to “ready to start.” Notion then invokes a Claude session. The agent draws context from connected pages, design systems, API documentation and product requirements documents before beginning the assignment. The task board supplies the handoff; the connected material supplies the working context.

Claude Managed Agents provides the long-running sessions and memory management behind that process, Liu said. The assignments are open-ended rather than limited to short answers: users can delegate coding, slides and spreadsheets without leaving Notion. The case study also lists PowerPoint presentations, PDFs and websites among the files agents can produce.

One engineering example begins with a product manager recording a meeting note. Notion transcribes it into action items, and Claude writes the code for a prototype. For nontechnical teams, the examples are client-facing work: presentations, brand strategy decks and sample websites. Liu described the division of labor as using Claude’s coding and file-generation strengths alongside Notion’s own agent.

The work need not proceed one assignment at a time. Notion says teams can start more than 30 agent tasks concurrently from one board, with outputs routed to the appropriate people for review. Liu illustrated that capacity with an account of starting 30 prototype tasks, leaving, and returning to find all of them completed. That example and the capacity figure are company-reported results.

Finished tasks feed the next session

The integration also links completed work to subsequent assignments. According to the case study, Claude identifies lessons from finished tasks and updates Notion’s shared skills database. Future sessions draw on those updated skills. Notion’s stated aim is for agent performance to improve over time without people having to maintain those instructions manually.

Notion’s next stated direction is to make familiar collaboration tools serve as the interface for agent work. The case study names task boards, suggested edits and version history; Liu also points to approval flows. He describes Notion as the “translation layer” between people and agents, using interactions teams already understand to organize their work together.

The company plans to add broader data sources, more connectors and deeper agent workflows spanning technical and nontechnical work. Those are future plans, distinct from the integration described in the case study. The next step Notion identifies is extending what agents can draw on and do within that collaborative workspace.