Writing
Notes from the field.
What I learn putting AI in front of people who do real work: how to measure it, how to make it safe to undo, how to get it used.
Who Will Know When the Agent Is Wrong?
The interface will dissolve into the work. What will not dissolve is the ability to tell when the agent is wrong, and we are about to automate away the very work that teaches it.
From Undo Rate to Useful Signal: Instrumenting Reversibility in Practice
Tracking reversions is easy. Interpreting them is hard. A user hitting 'undo' compresses multiple failure modes into a single binary event. Here is how to tease them apart.
Engineering notes
Longer write-ups that live with the code.
Build logs and design notes from my open-source projects, dead ends included.
- chessfpLEARNINGS.mdGitHub ↗Fingerprinting chess players from 5.7M moves: the training runs that collapsed, why, and what finally worked.
- codaINTERNALS.mdGitHub ↗How an agent harness with a first-class context namespace and line-level traces works inside.
- forgeREADMEGitHub ↗Git for agent runs: content-addressed, branchable, diffable, replayable.
- agent-feedback-uiREADMEGitHub ↗Four ways to put human feedback into an agent loop, and a test suite measuring how each changes the agent.