Writing
- 28% fewer input tokens
- 4% lower estimated cost
I cut 28% of my coding agent's input tokens. The cost fell about 4%.
A pre-registered experiment on trimming Codex's built-in instructions. Fewer tokens, almost the same cost, and worse results.
- 11.5 GiB peak memory before the fix
- 0.66 GiB after, same output
An AI-assisted PR is still your PR
I typed four sentences. AI agents found a quadratic memory bug in SPIRV-Tools, fixed it and opened the pull request. What was left for me.
- 1 bad decision by the agent
- 6 missing layers around it
One bad decision, six missing layers
An after-action review of the PocketOS database deletion. What the agent did, and the six infrastructure decisions that made it so hard to recover from.
Work
ArgonMongoDB sandboxes for AI agents. Open source, and I maintain it.
LineupFinderFestival lineups and set times. Run day to day by AI agents.
OpenPlainEnglishA browser extension that turns health jargon into plain English.
CHROMAVECTMusic. Otis Horizon is on Spotify.
Merged upstream
Bug fixes sent to open-source projects, working with AI coding agents.
