Key Takeaways

  • Off-the-shelf software leaves micro-gaps; custom AI bots succeed by filling narrow "software-shaped holes" that commercial products ignore.
  • Thompson ran Stratechery solo while handling writing, taxes, and customer support before spending years learning how to delegate to an executive assistant.
  • Big tech misunderstands consumer psychology by building AI shopping agents; consumers enjoy the discovery of shopping, but hate administrative chores like processing package returns.
  • Current AI agents simulate learning by writing state data to disk rather than updating model weights, creating operational limits for autonomy.

Custom Micro-Bots and Software-Shaped Holes

In the early days of Stratechery, Ben Thompson struggled under the weight of operational overhead. “I was doing all the writing, all the taxes, all the customer support, doing everything and basically dying,” Thompson recalled. Even after hiring a personal assistant, making delegation work required years of trial and error.

Today, Thompson uses lightweight custom tools like his home inventory system and his custom Gecko bot to handle tasks that commercial SaaS ignores. The goal is building targeted micro-tools that fill what Thompson calls "software-shaped holes" across daily workflows.

Consider Thompson's intake setup with his assistant. Over the weekend, instead of manually compiling tasks, Thompson snaps photos and sends quick notes directly to a bot. The system aggregates the inputs and automatically delivers an organized project briefing on Monday morning. “Just this simple concept of I tell the bot do this take a picture do this take a picture... he shows up Monday morning it sends a briefing here's all the stuff that Ben want over the weekend things like this ongoing projects his joy and like relief was unbelievable,” Thompson explained.

The Shopping Blindspot in Agentic Tech

Major tech platforms keep pitching agentic AI as an automated personal shopper or travel agent. Thompson views this focus as a basic misunderstanding of consumer behavior.

"People love shopping," Thompson noted. “Tech companies are like, number one, let us book your flight for you. It's like I've had a personal assistant for 10 years. He does not come within an inch of booking my flights for me. That is crazy talk.”

Consumers treat browsing, seat selection, and price comparison as leisure and personal control. They do not want an algorithm to eliminate an activity they enjoy. The real consumer value sits in post-purchase friction that people dread, such as managing package returns, disputing charges, or tracking repairs.

The Limits of Simulated Learning

When evaluating progress toward autonomous systems, Thompson uses a clear definition of intelligence: the ability to learn continuously.

“My personal what is AGI rule is AI that actually learns,” Thompson stated. “Right now what makes these agents possible is this like facsimile of learning where they're just writing stuff down all the time and that has lots of challenges.”

Because modern large language models cannot alter their weights during inference, developers rely on external memory files, context retrieval, and scratchpads. This approach works for simple routines, but it creates brittle state management as workflows expand.

What to Do With This

Map the repetitive administrative handoffs between you and your team this week. Write a simple script or bot that ingests raw mobile media and notes over the weekend, then auto-formats them into a prioritized Monday task briefing. If you are building consumer AI products, stop trying to automate the purchasing decisions users enjoy and focus on automating post-purchase return logistics.