Writing & field notes
Questions, measurements, and trade-offs from building with agents — newest first. Also available as an RSS feed.
Up to three models screened 10,336 test emails and repository files, mostly from public prompt-injection datasets on Hugging Face and GitHub. Email results look promising; repository results do not yet. An indication from a lab test, not production evidence.
A free, ten-page module on what a coding agent costs to run and why, built from four mechanics (re-reading, caching, expiry and generation) and ending with a controlled cost experiment you can run on your own work.
Four hypotheses tested in 40 measured agent runs across five Kubernetes operations (incident repair, a failed rollout, a capacity change, a memory fix and a Helm lifecycle), each evaluated per session with Agent Tokenwatch against independently verified success.