Notion’s Token Town: 5 Rebuilds, 100+ Tools, MCP vs CLIs and the Software Factory Future — Simon Last & Sarah Sachs of Notion

Simon Last and Sarah Sachs of Notion join the Latent Space podcast to discuss the evolution of their AI capabilities, particularly the launch of Custom Agents and the broader vision of a 'software factory' future. They reflect on five major rebuilds since 2022, highlighting how early attempts at agent development failed due to immature models, short context windows, and poor tooling—until breakthroughs in reasoning models and tool calling (around Sonnet 3.6) enabled real progress. The team emphasizes a culture of rapid iteration, 'demos over memos,' and a 'Simon Vortex' of high-velocity prototyping, where engineers are empowered to build, delete, and rebuild without ego. Central to their approach is a focus on user journeys over flashy tools, with agents built to solve real workflows like email triaging, lease processing, and meeting note summarization. They also explore the tension between MCPs and CLIs, arguing for CLIs in deterministic, permissioned tasks and MCPs for broader tool access, while advocating for value-aligned pricing and a 'Notion's Last Exam' framework to evaluate frontier capabilities. The conversation reveals a company deeply committed to being the system of record for enterprise work, where agents are not replacements for humans but enablers of deeper collaboration and productivity. The episode underscores Notion’s unique position: not just a tool builder but a platform that abstracts complexity through primitives like databases, pages, and evals, enabling agents to compose workflows seamlessly. They discuss the importance of agent self-healing, internal tooling, and the rise of 'model behavior engineers'—a new role focused on understanding model behavior, not just coding. Despite the excitement around AI, the team remains grounded in practicality, prioritizing robustness, security, and enterprise readiness. They also touch on the future of AI, including the potential for fine-tuning on company-specific data and the critical role of retrieval and ranking systems optimized for agent use. Ultimately, the vision is clear: Notion is not building the next AI model, but the best environment where agents can thrive and humans can focus on what matters most.