The $149.25 release candidate

Simon Willison's sqlite-utils is one of those quietly load-bearing developer tools: 12,000 GitHub stars, on every Python data hacker's machine, backing his own Datasette and llm projects. On July 7 it hit 4.0, its first major bump since November 2020. The headline feature is schema migrations landing natively in the library — sequential, decorator-defined Python functions tracked in a table so a database upgrades itself in order. Useful, but hardly new; Django and Rails have done this for a decade, and Willison's own standalone package did it for years (he archived it the same day).

The story is how the release got built. Willison titled the release-candidate post with the exact figure: mostly written by Claude Fable, for about $149.25 — 37 prompts, 34 commits, roughly 1,300 lines changed across 30 files, over a few days. He reached for the agent partly because he had limited time on his subscription's model. A concrete, dollar-costed data point on a real maintained library is rare; most agentic-coding claims are demos on toy repos.

Then the part that changed his mind. Willison had thought using one model to review another's work was 'somewhat absurd.' He handed the code to GPT-5.5, which found a genuine correctness bug: an UPDATE issued through the wrong method committed the write to the database before raising an error — the kind of silent data mutation that hides for months. He still did the final human review himself through GitHub's pull-request interface, and said it sharpened his own grasp of how SQLite handles transactions.

The migrations will matter to the people who use the tool. The release economics are the part with reach: a legible, verifiable receipt for what shipping real software with an agent — plus a second model as adversarial reviewer — actually costs and catches today.