How Notion Built Agent Orchestration on Claude to Cut Costs 90%

Notion started as a workspace where knowledge lives: meeting notes, product specs, process guides, and project docs. As the product grew to serve millions of users including major enterprises, the first AI challenge was making all of that knowledge findable. Customer support teams needed troubleshooting steps on demand. Sales reps needed to find process documentation without waiting. Product designers needed brand guidelines without searching manually. Claude became the engine for Notion’s Enterprise Search, AI Writer, and Autofill features, helping users query their entire workspace in natural language and get answers drawn from connected apps and internal documents.

But as AI agents became capable of producing real work, a second and more ambitious challenge emerged: most agent interactions were one-to-one, a single person with a single agent on a single machine. There was no shared visibility, no approval workflows, no way for colleagues to step in and review or iterate together. Notion wanted to bring agents into its collaborative model. As Product Manager Eric Liu described it: powerful agents were being built for every vertical slice of work — why not bring them all into Notion?

Notion selected Claude Opus for its agent layer after testing multiple providers. Co-founder Simon Last cited Claude’s response quality and instruction-following as decisive, especially for use cases where tone and feel matter. AI Lead Engineer Sarah Sachs highlighted Opus 4.6’s ability to interpret intent accurately and produce shareable outputs on the first attempt. Claude Managed Agents provided the infrastructure for long-running sessions, memory management, and high-quality outputs over time — without requiring Notion to build a custom agent runtime.

The result is a workflow that turns Notion’s existing task boards into an agent dispatch system. A team creates a task, moves it to “ready to start,” and Notion invokes a Claude session that picks up context from connected pages, API docs, design systems, and product requirements. For engineering teams, this produces prototypes and code changes. For non-technical teams, it generates presentations, brand strategy decks, and sample websites. Liu described kicking off 30 prototype tasks at once and returning to find them all completed. On the infrastructure side, prompt caching cut Notion’s costs by 90% and latency by up to 85% across the millions of daily AI interactions.

The measurable enterprise impact has been significant. Osaka Gas reduced time spent searching for information by 35%. Remote saves an estimated 10 minutes per search across 300 daily queries. dbt Labs eliminated the need for separate AI tools, saving over $35,000 annually. For new employees, the AI assistant is used 10-20 times daily in their first weeks. Notion is now building toward a workspace where the same interfaces humans use for collaboration — task boards, suggested edits, version history — also serve as the interface for directing agent work.