datasette code-frequency chart on GitHub

Simon Willison has been running a low-key experiment on himself: does using frontier coding agents actually change how much code he ships? Rather than guess, he went looking for hard evidence and landed on something almost embarrassingly simple — GitHub's built-in code-frequency chart for his own open source project, Datasette.

The chart plots additions and deletions over time, and Willison says it's the clearest signal he's found so far. There's a pronounced spike near the end of the timeline, and he ties it directly to a cluster of new model releases: Opus 4.8, GPT-5.5, something he calls Fable 5, and GPT-5.6 Sol. These are the tools he's been leaning on for agentic coding work, and the jump in commit activity lines up almost exactly with when he started using them.

What makes this interesting isn't the chart itself — GitHub has offered code-frequency graphs for years, and plenty of maintainers ignore them. It's that Willison, who has spent much of 2024 and 2025 writing publicly about what these coding agents can and can't do, is using his own repository as a control group. Datasette isn't a toy project; it's a mature tool with years of history, so a visible spike against that baseline means something rather than nothing.

He's careful to frame this as illustrative rather than conclusive. One chart, one maintainer, one project. But it fits the broader pattern he and others have been describing anecdotally: newer, more capable coding models aren't just marginally faster at autocomplete-style suggestions, they're changing how much a single developer can push through in a given week. Willison's instinct to go find receipts, rather than just assert the vibe, is very on-brand for him.