What an accepted task costs
A month of agent build tasks priced by the task a reviewer accepted, and where a 3.6 times gap between two models came from.
Essays on running AI agents in production, each built on something I measured in my own workspace, and a digest of the public work as it ships. New pieces also go out on Substack, where you can get them by email.
Latest
A month of agent build tasks priced by the task a reviewer accepted, and where a 3.6 times gap between two models came from.
Essays
When AI does the work, the people checking it stop understanding it, and the research says a training course won’t fix that.
Here's a cheap way to test an AI change without a research team.
The most reliable control in my agent workspace is a script that can't be persuaded, because it never reads the argument.
The same file read carried 147x more re-reading depending on when it loaded. That makes it an architecture decision.
Nothing is watching them. Four questions that need mechanical answers before the volume arrives.
An agent workspace runs while you sleep, and rots while you sleep. So mine audits itself about once a week. The hard part wasn't building the audit. It was the three ways a self-audit quietly lies to you.
How to put a stateless enterprise LLM assistant inside a workspace that gives it the persistence, provenance, and validation it doesn't have on its own.
Turning a scheduled agent from a question-asker into a sandboxed teammate that ships work for review.
Digest