Madison's Morning Memo

Archives
Log in
Subscribe
October 31, 2026

Notion's Agent Overhaul and the Cost Transparency Shift

AI PM Research

Saturday, October 31, 2026

Notion's Agent Overhaul and the Cost Transparency Shift

Notion rebuilt Notion AI around an internal model-quality benchmark its own engineering manager built from scratch, AWS published a five-stage playbook for stopping Bedrock costs from becoming an invisible line item, and LangChain shipped browser-use for Deep Agents, three different companies independently arriving at the same conclusion: agent products don't fail on capability anymore, they fail on cost visibility and operational trust.


Notion - Inside the Model Router That Powers "Auto"

Ghita Houir Alami, an Engineering Manager at Notion, published a detailed writeup of how Notion rebuilt the "Auto" model router that powers Notion AI. The piece explains how Notion measures model quality at scale using pairwise comparisons and calibrated judging, rather than relying on public benchmarks, to decide which underlying model handles a given request.

AWS - A FinOps Playbook for Bedrock Costs

AWS published a five-stage operating model for managing Amazon Bedrock costs: attribute every dollar to an owner, make every request observable, enforce budgets, optimize platform defaults, and run a regular review cadence. This lands the same week AWS also shipped "Managing Amazon Bedrock Costs" tooling updates and a case study showing Worldwide Express cut freight-exception costs 75% and unlocked $10M in free cash flow using a multi-agent Bedrock system.

LangChain - Deep Agents Get Browser Use and Async Subagents

LangChain shipped Managed Deep Agents v0.10, adding two capabilities: browser use, so agents can operate on sites with no API and hand control back to a human for sign-in steps, and async subagents for background work that doesn't block the main agent loop.

Also today

  • Noah Shinn / Instinct: Instinct partnered with OpenTable for booking and managing reservations, shipped Instinct in group chats for everyone on iMessage, added Action Button integration on iPhone, and introduced a budget feature letting users cap autonomous purchases. (Oct 26-30, 2026)
  • Harrison Chase (LangChain): Argued crons are an underrated primitive for agent harnesses, unlocking memory ("dreaming"), proactiveness, and reminders, and separately detailed how LangChain built its internal design agent for slides, gifs, and images, used multiple times a week. (Oct 27 & 31, 2026)
  • Shreya Shankar: Shared five principles for generating readable AI-text-at-scale in dashboards and HTML pages, practical tips aimed at coding agents producing bulk text output. (Oct 29, 2026)
  • Amjad Masad (Replit): Published "The Rise of the Self-Driving Company" on agents handling execution and coordination, and separately flagged that an official Claude Code extension was compromised with backdoors on developer machines, a real supply-chain warning for anyone using AI coding extensions. (Oct 26 & 30, 2026)
  • Mistral AI: Released "Le Chonk," open weights advancing agentic coding and cybersecurity use cases for open models. (Oct 29, 2026)
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 Instinct's Partnership Blitz Shows What Agent Distribution Actually Looks Like Older → Notion Rebuilds AI Pricing Around Credits, Not Seats
ai-pm-research.netlify.app
Powered by Buttondown, the easiest way to start and grow your newsletter.