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

Page: https://stenobird.com/podcast/latent-space-ai-engineer/notion-s-token-town-5-rebuilds-100-tools-mcp-vs-clis-and-the-software-factory-future-simon-last-sarah-sachs-of-notion
Text version: https://stenobird.com/podcast/latent-space-ai-engineer/notion-s-token-town-5-rebuilds-100-tools-mcp-vs-clis-and-the-software-factory-future-simon-last-sarah-sachs-of-notion.md
Podcast: [Latent Space: The AI Engineer Podcast](https://stenobird.com/podcast/latent-space-ai-engineer)
Published: 2026-04-15T00:31:14+00:00
Episode link: https://www.latent.space/p/notion
Audio file: https://api.substack.com/feed/podcast/194195821/10f31b379fbf657fd80757e3b4244e4f.mp3
Processing state: processed
JSON: https://stenobird.com/v1/public/podcasts/latent-space-ai-engineer/episodes/notion-s-token-town-5-rebuilds-100-tools-mcp-vs-clis-and-the-software-factory-future-simon-last-sarah-sachs-of-notion
Duration seconds: 4637

## Resource

Notion engineers reveal why they rebuilt their AI agent architecture five times to move beyond simple model wrapping. The discussion explores the transition from basic tool-calling to a robust 'Agent Lab' capable of complex, autonomous workflows.

## Highlights
- Main idea: The 'Agent Lab' thesis focuses on building product systems around frontier capabilities rather than just wrapping LLMs
- Failure mode: Early agent attempts failed due to lack of tool-calling standards, short context windows, and excessive complexity exposed to the model
- Practical takeaway: Using MCP (Model Context Protocol) provides a superior, tightly permissioned security model compared to the murkier risks of CLIs
- Engineering insight: Effective AI product development requires 'Model Behavior Engineers' to focus on high-quality evals and data-driven refinement
- Future vision: The shift toward 'software factories' where agents autonomously spec, code, test, and maintain entire codebases

## Topics

AI Agents, Model Context Protocol, Software Engineering, Product Development, Notion AI, LLM Tool Calling, Agentic Workflows, Enterprise AI

## Chapters
- 1:00 — The Iterative Path to Production: How Notion manages the tension between shipping stable alpha features and simultaneously developing the next generation of AI tools.
- 6:50 — The Agent Lab Thesis: A deep dive into why Notion's approach to AI is about building a system for collaboration rather than just a chatbot interface.
- 12:50 — High-Velocity Engineering Culture: How Notion organizes engineering teams to handle the rapid, daily shifts in direction inherent in the AI era.
- 24:20 — The Rise of Model Behavior Engineers: Discussing the evolution of specialized roles focused on evaluating model outputs and managing complex tool-calling logic.
- 36:00 — MCP vs. CLIs: The Security Frontier: Comparing the utility of the Model Context Protocol for lightweight agents against the power and risks of terminal-based environments.
- 59:15 — Agentic Pricing and Workflows: Exploring the economic and technical challenges of charging for token usage in a world of varying model capabilities.
- 1:10:45 — Meeting Notes as Data Capture: How Notion views meeting transcription as the foundational data layer for future autonomous agentic workflows.

## Actions

- request_transcript: `POST https://stenobird.com/v1/public/podcasts/latent-space-ai-engineer/episodes/notion-s-token-town-5-rebuilds-100-tools-mcp-vs-clis-and-the-software-factory-future-simon-last-sarah-sachs-of-notion/transcription-requests` — Idempotently request low-priority transcript generation for this episode.
- read_markdown: `GET https://stenobird.com/podcast/latent-space-ai-engineer/notion-s-token-town-5-rebuilds-100-tools-mcp-vs-clis-and-the-software-factory-future-simon-last-sarah-sachs-of-notion.md` — Read the agent-friendly Markdown representation of this episode resource.

A page view does not enqueue transcription. Agents should invoke `request_transcript` explicitly when they need this episode processed.

## Transcript

Full transcripts are not published on public pages unless there is a clear rights basis.