Why Notion AI Agent V3 Ditched Complex Code for Plain English
Discover how Notion AI Agent V3 achieved a massive performance boost by replacing complex data modeling with plain text and markdown. Insights from AI Lead Ryan Nystrom.
Complexity was the enemy of progress. When Notion first started experimenting with LLMs, they went heavy on advanced code generation and rigid schemas. It didn't work. Now, the team has pivoted to a startlingly simple approach: talking to AI like a human. Ryan Nystrom, Notion's AI engineering lead, says this "rewiring" is exactly what led to their most successful AI launch ever.
How Notion Rewired its AI Agent Architecture
Engineers are trained to love deterministic systems, but LLMs thrive on intuition. Nystrom's team abandoned "pretty complicated rendering" like JSON or XML in favor of Markdown. By representing Notion pages as plain text, they allowed the model to read, search, and edit content just as a person would. The results were immediate. Released in September 2025, Notion V3 and its customizable AI agents have seen usage patterns that Nystrom calls a "step function improvement."
The 150,000 Token Sweet Spot
More data isn't always better. The team discovered that stuffing the context window leads to a "slippery slope" of latency and inaccuracy. They found the golden ratio at 100,000 to 150,000 tokens. Beyond this, the model's performance degrades significantly. To keep the AI sharp, they also limited the "menu" of tools available to the agents, preventing the paradox of choice that often confuses complex models.
This content is AI-generated based on source articles. While we strive for accuracy, errors may occur. We recommend verifying with the original source.
Related Articles
Databricks CTO Matei Zaharia just won computing's top prize. His take on AGI, the security nightmare hiding inside AI agents, and why the real AI revolution is about research, not chatbots.
A surprise leak of Anthropic's Claude Code source code revealed 'Kairos'—a dormant background AI agent designed to act before you even ask. Here's what it means.
At GTC 2026, Nvidia is expected to unveil an inference chip and the NemoClaw AI agent platform. What happens when the company that owns 80% of AI training comes for the rest of the stack?
Rox just hit a $1.2B valuation on $8M ARR. The AI sales agent startup thinks it can replace a dozen fragmented tools—and maybe a few salespeople. Here's what's actually at stake.
Databricks CTO Matei Zaharia just won computing's top prize. His take on AGI, the security nightmare hiding inside AI agents, and why the real AI revolution is about research, not chatbots.
A surprise leak of Anthropic's Claude Code source code revealed 'Kairos'—a dormant background AI agent designed to act before you even ask. Here's what it means.
At GTC 2026, Nvidia is expected to unveil an inference chip and the NemoClaw AI agent platform. What happens when the company that owns 80% of AI training comes for the rest of the stack?
Rox just hit a $1.2B valuation on $8M ARR. The AI sales agent startup thinks it can replace a dozen fragmented tools—and maybe a few salespeople. Here's what's actually at stake.
Thoughts
Share your thoughts on this article
Sign in to join the conversation