Writing on the human side of agentic coding.
Field notes from the work: cognitive science, AI tooling, engineering management, the craft of writing software in a year where most of it is being written for us. Slow, considered, occasionally wrong, never recycled.
Recent writing.
Why "lines accepted" is the wrong metric for your engineering org.
Acceptance rate is a measure of workflow interruption, not productivity. It rewards the behaviour that produces risk and punishes the behaviour you want. Here's what to track instead: solo throughput trend, concept retention rate, and incident attribution by code authorship.
FSRS for code: why retrieval beats re-reading for engineers.
A walk through the cognitive science of spaced retrieval, why we picked FSRS over SM-2 for the scheduler, and how prompts get extracted from your own past commits.
The first 90 days of a junior engineer in 2026.
What we hear from EMs onboarding the cohort that joined the workforce after the AI tooling wave. The patterns are surprisingly consistent — and surprisingly fixable.
Building a concept tagger that doesn't lie.
How we annotate diffs with concept tags using a tree-sitter rule layer with an LLM fallback, and why the calibration of the confidence score matters more than the accuracy.
Bring-your-own-model isn't a launch limitation. It's a permanent stance.
The structural conflict between AI-coding companies' margin and your token bill, and why we will never become a token reseller — even if it makes us slower to grow.
"We caught it during the layoffs": an EM on bus-factor warnings.
A conversation with the engineering manager of a Series C company about discovering, mid-restructure, that three production systems were held in the head of one person.