AI-Assisted Development
Day-to-day engineering with a model in the loop.
Using AI across the normal work of building software: drafting code, writing tests, explaining unfamiliar code, reviewing diffs, debugging. Incremental rather than a new paradigm.
The gains come from the boring parts — boilerplate, tests, migrations, first drafts, searching a codebase you do not know. The parts that need judgment still need judgment, and review does not disappear.
The evidence is genuinely mixed and worth holding both halves. A controlled experiment on a specific task found a large speed-up; a randomised trial on experienced maintainers working in their own repositories found tasks took longer. Both results are real, which means the answer depends on task shape, familiarity and how the tool is used.
Where the time actually goes
A team wants to know whether their tooling is paying off.
Measure your own workflow rather than citing someone else’s study. The two most common findings are that generation is fast and verification is not, and that gains concentrate in unfamiliar code and boilerplate while experienced work in familiar code can get slower.
Reading productivity studies honestly
A headline number is quoted in a planning meeting.
Check the population, the task and the year. A single narrow task, or a tool set from eighteen months ago, does not transfer to your team. The research is useful for the mechanisms it identifies, not as a forecast of your results.
- paperThe impact of AI on developer productivity: evidence from GitHub Copilot ↗
A controlled experiment: developers with Copilot finished a specific JavaScript HTTP-server task 55.8% faster. One narrow task — do not generalise the number to software work.
- reportMeasuring the impact of early-2025 AI on experienced open-source developer productivity ↗
A randomised trial on experienced maintainers’ own repositories in which tasks took 19% longer with the AI tools, against participants’ expectation of a speed-up. A snapshot of one population and tool set.
- reportWe are changing our developer productivity experiment design ↗
METR explains why self-selection in their follow-up experiment made the productivity estimate unreliable. Read alongside the 2025 result before quoting either as a verdict.
- reportState of AI-assisted software development 2025 ↗
Frames AI as an amplifier of the organisation it lands in: the same tooling helps strong delivery cultures and compounds problems in weak ones.
Vibe Coding
Building by prompt-and-accept, without reading the code.
Human Review
A person reading the work, with judgment as part of the job.
DX · Developer Experience
How it feels to build with a tool or platform.
Harness · Scaffold
Everything around the model that makes it an agent.