The Notion AI agent fundamentally changes how users interact with the platform, shifting from manual creation of documents and databases to a conversational interface. This abstracts away the complexity of Notion's powerful features, making it accessible to a broader audience who can now accomplish tasks through natural language commands.
Notion explicitly defines itself as an "applied AI company," choosing to integrate best-in-class third-party models rather than building foundational models in-house. Their focus is on the application layer, including prompt engineering, data representation, and building a robust tool-use framework.
A critical technical insight was the decision to represent Notion's complex block-based structure as simple Markdown for the LLM. This concept of an "Agent Computer Interface" (ACI) — the layer between the AI and the software's internal data structures — was key to achieving reliability and performance.
Notion employs a sophisticated, two-tiered evaluation process for its AI agent. It uses a "golden set" of tests that must achieve 100% accuracy for core functions, and a separate, more challenging set designed to measure incremental improvements from new models or techniques.
Notion is developing customizable, shareable, and autonomous agents that can be triggered by events. Internal use of these custom agents led to a quadrupling of AI usage, indicating a significant shift in how workflows can be automated and managed within teams.
Keep pulling the thread on Akshay Kothari & Ryan Nystrom.