Enterprise Knowledge Search · IT Operations Automation
Notion needed to embed trusted Claude agents natively inside collaborative workflows, with automations, long-running tasks, memory, coding sandboxes, and human approvals that could scale across whole workspaces instead of one-user/one-agent chats.
Reported impact
Notion created 18,000 Claude agents in the first three weeks, saw 90 percent of agent turns come from automations rather than chat, and reduced early prototyping work from about 12 hours to 20 minutes while keeping agent collaboration and permissions inside the existing workspace.
Notion built External Agents on Claude Managed Agents so Claude sessions can live inside Notion pages and task boards rather than separate tools. A Notion task board acts as the orchestrator: moving a task to ready-to-start invokes a Claude session that pulls context from connected pages, design systems, API docs, and product requirements documents. Most usage is automated, triggered by Notion pages, Slack messages, or schedules, and top workspaces run several agents across dozens of threads per day. Managed Agents supplies the long-running execution layer, memory, sandbox, packages, and GitHub endpoint that let Claude read repositories and run code. Notion also splits work across stage-specific agents—investigate, plan, build—with different permissions and sharing rules so collaborative teams can review and approve work safely inside the same workspace.
Pages, task boards, Slack messages, or schedules trigger a Claude session inside Notion; the session pulls context from connected documents and repositories, performs the requested work, and streams results back as files, comments, or page updates that teammates can review and approve.
90% of the turns people are doing with Claude are automations in Notion, triggered on a Notion page, a Slack message, or scheduled runs.
Even with the early prototype, we saw 12 hours of prototyping work collapse into about 20 minutes.
So far it's about 18,000 Claude agents created, and those agents have done about 140,000 steps.
The GitHub endpoint in Managed Agents was really useful.
You create a task, move it to "ready to start," and it invokes a Claude session.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Anthropic Sample Archived · Aug 30, 2026 Atlassian Atlassian wanted agents to do consequential work inside Jira, Confluence, and related products, but enterprise customers would only trust long-running automation if the models could follow complex instructions, use… 5 million+ agents executed in customer business workflows each month Claude Code · Claude Platform · Rovo Chat Anthropic Sample Archived · Apr 16, 2026 Slack Slack needed to turn huge volumes of messages, threads, and files into searchable, permission-aware knowledge and time-saving summaries without sacrificing security or conversational quality. 97 minutes per week saved by the average user through summarization and recap features Claude Code · Claude Platform Anthropic Sample Archived · Feb 10, 2026 HubSpot HubSpot's engineering, marketing, and customer success teams were spending too much time on migrations, maintenance, content creation, troubleshooting, and personalized follow-ups that did not scale across the… From 3-5 days to under an hour Claude Code · Claude Platform · Claude Sonnet 4.5 