Ten months of Android, two of them with an agent

I looked back at 10 months of my Android work: about 580 commits and 106 pull requests. For the last two months, an AI coding agent (Claude Code) wrote most of the code while I planned, reviewed and shipped.

Speed

About 4× more pull requests per month, each about 3× larger. Most of it was a design-system migration across a dozen features, plus a deeplink test harness with CI.

Quality

Reviewers left 63% fewer comments per line changed, and the comments that remained were about behavior, not style.

Calibration

I analyzed my own prompts. The times I had to interrupt the agent fell from 10.7 to 0.9 per 100 prompts. Corrections stayed flat, but most of them are now new requirements from QA and design, not misunderstandings.

The risk I'm watching: knowledge stagnation

When the agent explores the code, writes the fix and drafts the review reply, I learn less by default. The deep skills I rely on (lifecycles, concurrency, release mechanics) came from the eight months before.

So I'm tracking a few signals: whether I can explain a merged change without opening it, and doing small tasks by hand in areas the agent usually handles.

The speedup is real. Keeping my own understanding current takes deliberate effort.