Notion's latest AI breakthrough, customizable AI agents, stemmed from a surprising source: simplification. Released in September as part of V3 of its productivity software, the feature's success is attributed to a pivot in the company's AI development strategy.

Initially, engineers at Notion AI experimented with advanced code generation, complex schemas, and heavy instructioning when working with large language models (LLMs) and agentic AI. However, according to Ryan Nystrom, Notion AI engineering lead, trial and error revealed that eliminating complicated data modeling dramatically improved model performance. Nystrom and his team shifted to simple prompts, human-readable representations, minimal abstraction, and familiar markdown formats.

This re-wired approach led to the creation of customizable AI agents, which have become Notion's most successful AI tool to date. Nystrom described the improvement as a "step function," based on usage patterns compared to previous versions. "It's that feeling of when the product is being pulled out of you rather than you trying to push," Nystrom explained in a VB Beyond the Pilot podcast. "We knew from that moment, really early on, that we had something. Now it's, How could I ever use Notion without this feature?"

The development highlights a crucial lesson in the evolution of AI: sometimes, less is more. The initial instinct in AI development often leans towards complexity, assuming that more intricate models and instructions will yield better results. However, Notion's experience suggests that focusing on simplicity and human readability can unlock greater potential. This approach aligns with the growing understanding that LLMs, while powerful, can be sensitive to the nuances of input data. Clear, concise prompts and data structures can minimize ambiguity and improve the accuracy and efficiency of AI models.

The implications of this approach extend beyond Notion. As AI becomes increasingly integrated into various aspects of society, the ability to create accessible and user-friendly AI tools becomes paramount. Simplifying AI development processes can empower a wider range of individuals and organizations to leverage the technology, fostering innovation and addressing diverse needs. Moreover, focusing on human-readable representations can enhance transparency and trust in AI systems, making them more understandable and accountable.

The success of Notion's AI agents underscores the importance of iterative experimentation and a willingness to adapt development strategies based on real-world feedback. As AI technology continues to evolve, companies must remain agile and open to new approaches, even if they challenge conventional wisdom. The next phase for Notion involves further refining its AI agents based on user feedback and exploring new applications for the technology within its productivity platform.