Works On My Machine logo ~/works-on-my-machine

❯ Notes from inside developer platforms. IDP, CI/CD, DX, and the gaps between them.

2026-11-17 18 min #agents#git#developer-productivity

One Worktree, One Issue, One Agent

I ran a four-week worktree experiment for five months. It didn't cut my context switching. It let several agents run at once, and every new worktree quietly lost its untracked config. The loop, the hook that fixes it, and why I went back to a plain clone.

2026-10-27 12 min #ai#agents#context-engineering

Three Ways to Go Broke on Context

Hand an agent two conflicting requirements and it may not stop — mine split the difference. The three kinds of missing context, the matching defense for each, and why the defense has to be cheap to carry.

2026-10-12 7 min #ai#agents#code-review

The Guess Comes Before the Diff

Part 1 built the review machine that keeps agent-written code honest. But verification keeps the code safe — it doesn't keep me sharp. A Craft Conference talk named the hole in my pipeline, and the fix costs three minutes per PR.

2026-10-07 10 min #ai#agents#claude-code

The Expensive Part Is Remembering

A long agent session cost me $6.89, and the model wrote almost none of it. Two-thirds of the bill was the agent re-reading its own past — and the fix was architectural, not stylistic.

2026-09-28 6 min #ai#agents#developer-productivity

Pencils Down Is a Review Process, Not a Vibe

DHH declared 37signals pencils down at Rails World 2026. But agents writing your code is the easy part. The policy only works if you can answer: who reviews, against what bar, and what happens when the review fails? Here's the review machine I built to make pencils down survivable.

2026-07-16 10 min #ai#agents#github-copilot

The Implicit Signals Your Cloud Agent Is Reading

The prompt is only one input. The cloud agent also reads your issue structure, your git hooks, your hook decisions, and your environment variables — most of it silently, and not always the way you'd expect.

2026-04-21 14 min #ai#agents#github-copilot

GitHub Cloud Agents Work When You Build the Feedback Loop

After 3-4 days of bad PRs, I learned the hard part of using a GitHub Cloud Agent isn't the prompt — it's the environment, hooks, and validation layers that tell the agent whether it actually succeeded.