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

The episode begins with a deep dive into Notion's journey with AI, highlighting how their custom agents feature was rebuilt multiple times—around four to five iterations—due to the rapidly changing landscape of AI capabilities. Early attempts in 2022 faced significant issues such as model shortfalls, lack of standard tool-calling protocols, and short context windows, making the technology unreliable for production. Major unlocks came with improvements in model versions (like Sonic 3.6/7) and the development of more reliable, background-running custom agents. The focus is on building systems that are edge-to-edge useful, like permissions management, where multiple layers of permission swings had to be refined to ensure safe operation. The overall theme emphasizes adapting to frontier capabilities by not fighting the model, but aligning product development with evolving AI limits and opportunities.

Notion's strategy is underpinned by the 'Agent Lab' thesis: developing AI solutions not just by wrapping models, but by understanding how humans collaborate and building systems that support this process. The discussion covers how early product development involved wrappers around functionalities, but the focus has shifted toward establishing primitives, primitives, like specs and testing frameworks, to build robust 'software factories.' The idea is to distribute tool ownership, incorporate self-verification loops, and enable agents to debug and update themselves—extending towards fully automated pipelines. This reflects a broader philosophy of engineering with agility, prioritizing capabilities, and building for sophisticated operators instead of general usability for all.

Sarah Sachs and Simon Last emphasize the importance of a culture of low ego, rapid prototyping, and continuous iteration—what they call 'demos over memos.' Hierarchical rigidity is avoided, enabling teams to rebuild and shed code quickly. Hackathons, internal 'Simon Vortex' projects, and early security involvement foster a culture of speed, curiosity, and safety. Sachs highlights her role in aligning teams around objectives rather than ideas, advocating for a culture that centers on quickly testing prototypes, learning from failures, and allowing individuals to pursue their own ideas. Leadership is about setting clear objectives and creating an environment where experimentation is embedded into daily routines.

The episode details how Notion’s internal architecture for AI tools has evolved over time. They started with a JavaScript-based coding engine, moved to an XML representation for tool calling, then to a Markdown-based approach—aiming for simplicity and model compatibility—and finally transitioned to a SQL-like database system (SQLite) for easier querying by models. Notion’s approach is to expose the model to the environment it prefers, simplifying interactions and reducing complexity. They manage tool composition through shared primitives, allowing agents to invoke other agents or write into databases, with a focus on minimal human intervention and maximum robustness. This evolution reflects an engineering mindset of reducing system complexity while increasing capability and flexibility.

Notion emphasizes using AI-powered tools to streamline workflows such as meeting summaries, email triaging, bug tracking, and tenant applications. They build live demos, like a coworking space application agent, that showcase how agents can automate routine tasks: parsing emails, enriching profiles via web search, and logging structured data. The discussion highlights how internal systems, like meeting notes, serve as vital data flywheels—improving search, task management, and collaboration. They also explore complex composition strategies—using manager agents to coordinate multiple sub-agents—evolving toward a seamless, self-healing, and permissioned multi-agent ecosystem that minimizes the need for manual setups.

Participants discuss the merits and drawbacks of using Protocols like MCP (Model Communication Protocol) versus CLI in Notion’s AI ecosystem. Sarah Sachs supports MCP for its simplicity, permissioning, and deterministic nature, especially suited for tightly permissioned, low-risk environments. Simon Last expresses enthusiasm for CLI—its power in the terminal environment, self-debugging, and ease of coding agent fixes—though it introduces complexity around security and token management. The conversation underscores the importance of selecting the right tool for the right context: MCP provides a straightforward, safe protocol, while CLI offers flexibility and depth for advanced users and research settings. Both can coexist as different layers of the stack, tailored to specific needs.

Pricing strategies are discussed in depth, focusing on the use of credits as an abstraction over token throughput. Notion aims for price fairness and transparency, avoiding complex, per-task pricing models due to implementation difficulty. They emphasize that different models and tiers—auto, reasoning, open source—should be matched to user tasks and latency requirements. Sachs explains their cautious stance on training custom foundation models, preferring to fine-tune for specific enterprise needs without retraining from scratch. The episode concludes with visions of the future: modular, context-aware, fine-tuned models. Notion aspires to be an adaptable, capable AI platform that supports dynamic model switching to balance capability, cost, and latency.