I'm Mike McMahon — infrastructure/SRE engineer, AI practitioner, homelab builder. This site is where I document the work: what I'm building, how it broke, and what I learned along the way.
Use the sidebar to navigate. Start with How I Got Here if you're new, or jump straight into the Session Notes for technical content.
Subscribe to new posts: Email digest (daily) or RSS feed
The newsletter on this site has never sent an email. Not one, in five and a half months. Pre-commit checks, a multi-browser CI suite, a merge gate, RPC-only database access and a cron secret all stayed green the whole time, because every one of them was pointed at something other than the thing that broke. The only alarm was a vendor telling me an SMTP key was about to expire.
Months after the first run, I re-ran the multi-agent eval across a fresh generation of models - Claude Fable 5, GPT-5.5, GPT-5.6 Sol. I expected the frontier tier to pull ahead. Instead the cheap model from a generation ago tied for first, one handwriting sample decided the entire ranking, paying 33× more bought nothing - and when I stress-tested that conclusion, half the quality signal turned out not to reproduce.
Building a Python CLI that projects cost across six models by task type and complexity, using real cost data from the agent eval harness and OCR eval.
Mid-task messages now reach the agent while it is still working, and it changes the shape of the session more than any capability upgrade has. Notes from a day where three wrong turns each cost one sentence to correct instead of a full re-brief, plus the honest limit: steering got better, correctness did not.