Episode Summary
Executive Summary: Simon Last explains how Notion’s AI strategy evolved from a fast-shipped writing assistant into a platform for humans and agents to collaborate. He argues the product now centers on indexing, retrieval, custom agents, and agent-friendly APIs, while the company’s engineering culture has shifted toward smaller teams, rapid rewrites, and agent-driven development workflows.
Main Topics: Notion’s AI origin story and GPT-4 inflection point (Priority: 5/5): Last describes seeing GPT-4 as the moment AI became real enough to anchor a product strategy, prompting immediate work on both short-term and long-term AI features. From writing assistant to general assistant (Priority: 5/5): Notion first shipped AI Writer for rewriting and drafting, then expanded to workspace Q&A, and later to agents that can create, edit, and query documents and databases. Indexing and retrieval as the core technical moat (Priority: 5/5): A major theme is that high-quality semantic indexing across Notion, Slack, Google Drive, and other sources is hard but essential, requiring empirical iteration and source-specific tuning. Agent-first product and API design (Priority: 5/5): Notion reworked its APIs and formats to be easier for agents than for humans, including a markdown dialect for pages and SQLite-like interaction for databases. How coding agents changed Notion’s engineering org (Priority: 4/5): Last says coding agents have increased individual output, made prototyping more chaotic and ambitious, and pushed engineers toward end-to-end task design and verification rather than manual coding. Custom agents as autonomous knowledge workers (Priority: 5/5): Custom agents can be assigned a role, granted scoped access, and run in the background to triage email, route bugs, or respond in Slack, with memory improving over time. Notion as a collaboration layer for humans and agents (Priority: 4/5): Last frames Notion’s broader mission as a workspace where humans manage agents, with the same primitives—documents, databases, boards—still essential but newly oriented toward agent orchestration.
Key Arguments: GPT-4 made AI product opportunities concrete because it was already smart, instruction-following, and knowledge-rich enough to justify immediate product work. The fastest successful AI launch was the writing assistant because it was a single-step task without retrieval complexity. Workspace search and Q&A were much harder because indexing and retrieval required real-time updates, evaluation, and source-specific tuning. Most companies are surprisingly poor at indexing their own data; winning here requires craft, attention to detail, and iterative testing. AI reduces the importance of human organizational structure in some cases; with embeddings, the model cares more about good text snippets than tree hierarchy. Engineering velocity and ambition rise sharply when agents can implement, verify, and maintain code end-to-end. The gap between average and top engineers widens when tools are used well; the best users can become dramatically more productive. Notion is designing for a new customer: the agent, not just the human user, which means APIs must be made agent-friendly. Custom agents can bootstrap their own capabilities over time, including potentially building integrations they don’t natively have. Notion’s strategic position is to be model-agnostic and let customers switch between frontier and open-source models as the landscape changes. The company still believes its core primitives—documents, databases, coordination structures—remain essential, but now for human-agent collaboration rather than only human work.
Data Points: GPT-4 discovery year: 2022 - Last says he and co-founder Ivan realized the opportunity after playing with GPT-4 at a company off-site in Mexico. AI Writer launch: February 2023 - Notion’s first AI-specific feature, focused on rewriting and drafting text inside documents. Workspace Q&A GA launch: October 2023 - Notion launched semantic workspace question-answering after a beta period. Personal agent launch: Last August or September - The full Notion personal agent shipped in the prior fall and can access the user’s permitted workspace data. Custom agents launch: Last week - At the time of the interview, custom agents had just launched and could be configured for autonomous background work. Harness rewrite cadence: Every 6 months - Last says Notion rewrites its AI harness roughly twice a year to keep up with model progress. Coding agents adoption start: Around April of last year - Last says he began using coding agents, including Claude Code, around that time. Personal coding-agent runtime record: 13 days straight - He describes one coding agent session running continuously for nearly two weeks. Email automation outcome: 95% - Last says 95% of his work and personal emails are not useful to him and are now filtered by an agent.
Pivotal Quotes: "the time is now to start, but think about how to apply this" — Simon Last: Describing the immediate reaction to GPT-4 after seeing both its intelligence and breadth of knowledge. "the goal is to create the best tool for humans to manage agents to do the work for them" — Simon Last: Explaining how Notion’s core product philosophy changed in the AI era. "I think of coding agents as the kernel of AGI" — Simon Last: Arguing that code is a powerful deterministic primitive and that coding agents are foundational to broader agent systems.
Implications: Notion is positioning itself as an agent orchestration layer, not just a productivity app. For builders, the message is that AI products must be redesigned around model capabilities, agent-friendly interfaces, and continuous evaluation—or risk becoming obsolete.