From Coder to Manager: Navigating the Shift to Agentic Engineering with Notion Co-Founder Simon Last

The Genesis of Notion AI: A Two-Fold Vision

Notion’s journey into AI began in earnest with early access to GPT-4 in 2022. The model’s ability to follow complex instructions and its vast world knowledge made it “instantly clear” that the time was right to integrate this technology. This led to a dual-pronged strategy:

1. Short-Term Vision: The Writing Assistant. The most immediate application was an in-document writing assistant. This feature, which allowed users to select text for rewriting or generate new text, was developed by a tiger team and launched within two to three months, in February 2023. It was a single-step task without complex retrieval, serving as a direct interface to the model’s capabilities.

2. Long-Term Vision: The General Assistant. The more ambitious goal was to create a general assistant or agent that could use all the tools available to a human user within Notion—creating databases, manipulating documents, and weaving them together to accomplish longer-range tasks. This vision took significantly more time to realize.

Technical Hurdles: Semantic Indexing and Retrieval

The path from a simple writing tool to a Q&A-capable assistant was a major technical undertaking, launched in October 2023. The core challenge was building a real-time, updating semantic index of a user’s entire workspace. This effort extended beyond Notion’s own data to external sources like Slack and Google Drive.

Simon Last notes that while many companies struggle with effective internal search, Notion succeeded through a combination of “AI pilled savviness” and meticulous “craft and attention to detail.” Key learnings included:

• Source-Specific Tuning: A one-size-fits-all approach to indexing and retrieval doesn’t work. Querying Slack is fundamentally different from querying Google Drive, and each data source requires a unique, empirical, and iterative approach to tuning.
• The Power of Embeddings: With semantic search, the user’s organizational structure (e.g., page hierarchy) becomes less critical. The AI primarily cares about finding the relevant snippet of text, regardless of where it’s stored. The advice now is to “not worry as much about organization” and focus on piping all relevant data into the system.
• Constant Iteration: The technical components, such as chunking strategy and retrieval pipelines, required significant iteration to perform well across diverse content bases.

The Agentic Shift in Engineering Culture

Notion’s internal development process has been radically transformed by AI, particularly by the use of coding agents.

• The Six-Month Rewrite Cycle: The AI harness—the system that orchestrates model interactions—is rewritten roughly every six months. This rapid iteration is crucial for deeply integrating the capabilities of the latest models, a practice many companies get wrong by sticking with a single implementation for too long. The use of coding agents makes this ambitious rewrite schedule feasible.
• A New Era of Development: The engineering team has moved from simple tab-autocomplete to fully-fledged coding agents that can implement, verify, and maintain features end-to-end. This has dramatically increased the ambition of what the team can build. As Simon states, “I haven’t written code since like last summer. I don’t type code anymore.” His role has shifted to being an “agent manager,” designing end-to-end tasks and verifying the results.
• Impact on Team Dynamics: The individual impact of an engineer has skyrocketed, creating a wider gap between the median and top performers. The culture has become more “messy and chaotic” in a positive sense, with a proliferation of high-fidelity prototypes. For instance, the design team built a “design playground”—a simplified Notion with UI primitives and an agent—to quickly spin up interactive prototypes instead of relying on static mocks.
• Rethinking Code Quality: While agents can produce large, complex pull requests, they also enable a higher standard of testing. It’s now standard to demand that agents produce fully end-to-end tested changes, leading to a higher degree of confidence in the code’s correctness. The human’s role shifts from writing code to carefully thinking about the change, its verification, and its safe deployment.

From Personal to Custom Autonomous Agents

Notion has progressively rolled out its agentic capabilities to users.

• Personal Agent: Launched in the fall of 2023, every Notion user has a personal agent with access to all the same information and tools the user has.
• Custom Agents: A more recent launch allows users to create custom agents for specific, autonomous tasks. These agents start with no permissions and must be explicitly granted access to data sources (like a specific database or a Slack channel). They can then run in the background to perform jobs like triaging feedback from a Slack channel into a task database.
• Bootstrapping Capabilities: The future vision is for agents to bootstrap their own capabilities. A key element of this is making coding a core tool for the agent. If an integration doesn’t exist, the agent should be able to build, deploy, and use it on the fly. This positions coding agents as “the kernel of AGI.”

Designing for a New User: The AI Agent

A significant challenge was realizing that Notion’s API, originally designed for human developers, was not suitable for LLMs. The agent is now treated as a “new customer.”

“We essentially have a new customer which is the agent.”

This required an empirical, iterative process of “user research where the user is actually the agent.” By observing where the agent struggled (e.g., with verbose JSON formats or excessive token usage), the team designed new, more convenient APIs:

• Pages: A custom Markdown dialect was created that extends standard Markdown to represent all of Notion’s blocks, which models are naturally good at parsing and generating.
• Databases: Agents interact with databases using SQLite, a language they are highly proficient in.

A New Paradigm: From a Tool for Work to a Tool for Agents

The core conception of Notion has fundamentally shifted.

• Before AI: “Create the best tool for humans to directly perform their work.”
• After AI: “Create the best tool for humans to manage agents to do the work for them.”

This represents a monumental change in product philosophy. However, the foundational primitives of Notion—documents for unstructured text, databases for structured data—remain incredibly relevant. They serve as the essential coordination layer for humans to manage their “swarm” of agents. A Kanban board, for example, is just as useful for tracking the work of 100 coding agents as it is for a team of 10 humans.

• Tags:
• AI Agents
• Software Development
• Large Language Models
• Feature
• Retrieval Augmented Generation
• Productivity
• Engineering Culture
• Notion