Agents · 1 min · May 16, 2026
Why Agent Reliability Is a Distribution Problem
Single-step accuracy looks fine. Then you chain ten steps and the product falls apart.
The notebook
27 dated entries, newest first. Teardowns, field notes, build logs, and signals.
Agents · 1 min · May 16, 2026
Single-step accuracy looks fine. Then you chain ten steps and the product falls apart.
Physical AI · 1 min · May 13, 2026
The race in robotics is not for better motors. It is for hours of demonstrated motion.
Takes · 1 min · May 11, 2026
A copilot suggests. An agent acts. Flattening the two is how teams get surprised by their tools.
Agents · 1 min · May 10, 2026
Coding agents quietly moved from autocomplete to running the loop. Here is what actually changed, and what it asks of the people who build.
Agents · 1 min · May 10, 2026
The failure mode that matters is not the agent that stops. It is the one that finishes, wrong.
Industry · 1 min · May 7, 2026
Every step down the cost curve quietly makes a new category of product economic.
Future · 1 min · May 4, 2026
When a tool costs an afternoon to build, the long tail of software gets very long.
Agents · 1 min · May 3, 2026
Spawning a subagent buys isolation and parallelism. It also fragments context. Pick deliberately.
Future · 1 min · May 2, 2026
Five concrete shifts I expect once agents are cheap, reliable, and everywhere — and one thing that will not change.
Physical AI · 1 min · Apr 30, 2026
The policy works in the simulator and fails in the room. The gap is where the engineering lives.
Takes · 1 min · Apr 27, 2026
A number that goes up is not the same as a product that got better. The gap is widening.
Agents · 1 min · Apr 25, 2026
Once the prefix is nearly free, the cheap move and the smart move stop being the same thing.