Papers

• Done is a column, not a fact 2026-10-08

  Every tool we track work in records what a human declared. That was fine when a human did the work. Now the code arrives faster than anyone can read it, and the tracker still takes our word for it.

• The business always wanted a date 2026-10-08

  Story points were a translation layer between how engineering works and how a business commits. Coding agents broke the layer, and what replaces it is closer to what the business asked for in the first place. The cost shows up somewhere less obvious.

• Team size has a ceiling, and agents lowered it 2026-10-08

  Coding agents broke the link between complexity and effort, so velocity stopped carrying information. What replaces it is a limit on how many contributors a piece of work can absorb, and the evidence says agents push that limit down rather than up.

• One developer, production scale 2026-09-14

  What working with AI coding agents actually buys a sole developer running a multi-tenant platform in production, and the three things it does not buy.

• Agents don't fail at writing code. They fail at knowing what already exists. 2026-09-14

  The hard part of agentic development is not generation. It is knowing what is already in the tree, and nothing about a language model makes it good at that.

• A gate that only ever passes is no gate 2026-08-28

  Most test suites prove that code works. Almost none prove that the tests would notice if it stopped.