Madison's Morning Memo

Archives
Log in
Subscribe
October 27, 2026

Notion's AI Credit System and LangChain's Internal Agent Playbook

AI PM Research

Tuesday, October 27, 2026

Notion's AI Credit System and LangChain's Internal Agent Playbook

Notion rebuilt its entire AI pricing model around a credit system while LangChain published the actual internal architecture of the agents running its marketing, sales, and engineering teams, and both are answering the same question from different directions: how do you make AI agents cheap enough and trustworthy enough to run constantly, not just on-demand.


Notion - Rebuilt Notion AI Around Credits, Model Routing, and Long-Running Agents

Notion rebuilt its AI product around three changes at once: a smarter model router that decides which model handles each request, long-running agents that can work in the background, and a bring-your-own-AI option that lets customers plug in their own model keys. The whole thing now runs on a credit system instead of flat per-seat AI pricing.

LangChain - Published the Five-Part Architecture Behind Its Internal Company Agents

LangChain documented the architecture pattern behind the agents it built for its own marketing, sales, and engineering teams, what Harrison Chase has been calling a "company harness" on X this week. The pattern has five parts: agents live inside Slack rather than a separate app, each agent handles identity per-tool rather than one shared credential, they keep both global and per-user memory, they execute code in a sandbox, and most of them run on a schedule rather than waiting to be invoked.

OpenAI - Published Usage Data From 165 Work Activities Across Nearly 100 Countries

OpenAI released a data-driven look at how ChatGPT usage breaks down across work activities and geographies, tracking 165 distinct work activities in nearly 100 countries and how usage patterns shift over time. This is the kind of longitudinal usage data that's rare to get from a foundation model company directly, most of what we see is anecdote or isolated case study.

Also today

  • Lenny Rachitsky: Published a "How I AI" episode on Reddit's custom Figma MCP + Cursor pipeline, built by principal UX engineer Blair Metcalf, that takes designers from blank slate to a shareable, production-shaped prototype
  • Teresa Torres: New All Things Product episode with Petra Wille on "Defining What Good Looks Like," framework-focused discussion on setting product quality bars
  • Pieter Levels: Spent a full day having Claude Code run a security audit across every VPS he operates, a concrete look at using coding agents for infrastructure security work rather than feature-building
  • OpenAI: Introduced GPT-6.1 Terra, bringing "fast and capable computer use to everyone," now generally available on Amazon Bedrock and AI Gateway same-day
  • Simon Willison: Published a hands-on evaluation of GPT-6.1 Terra, testing its computer-use capabilities directly
Read the full update →

AI PM Research · Daily updates on AI products and strategy

Don't miss what's next. Subscribe to Madison's Morning Memo:
← Newer Lenny's Personal Agent Showdown and GitHub's Agent-Native Developer Data Older → OpenAI's Work Data, Notion's Agent Credits, and GPT-6.1 Terra
ai-pm-research.netlify.app
Powered by Buttondown, the easiest way to start and grow your newsletter.