Writing

I put Claude Code in a cronjob in a k8s namespace. It is tasked to monitor an application and in the unfortunate case of application error (degraded pods), it is tasked to do a hotfix and document it.

A Claude Code instance runs without any human action for 24-hours. I gave it short-term (sqlite3) & long-term (qdrant) memory, as well as access to a browser.

I attempted to create a fully-autonomous Claude Code. While I did not reach full autonomy, I got pretty close. In this piece I share the stats, how it went, and well.. how it made me feel.

Splitting a column on a live table with 200k rows in it, using expand/contract, with every phase rehearsed on a Neon branch first. 319,568 requests went through the app while it ran and none of them failed.

An agent changes your database and you find out later. git diff covers the migration file, not what the migration did, and nothing at all about the row it updated on the way past. So I built a TUI you leave open while the work happens, on a branch the agent can wreck, and made it mark the numbers it cannot prove.

Databricks bought Neon and turned it into Lakebase, so an agent can get its own branch of production in seconds and the rows it writes land in a governed table you can query. I walk through what that gives you, where it stops (nothing merges back), and what it means for getting an agent past read-only.

Denis ran the same task through ten isolated Claude Code agents at once — each with its own git worktree and its own Postgres branch — and diffed what all ten did to the database. They agreed completely on the schema, then split 5/5 on two decisions nobody asked them to make. Where they split is the cheapest eval you will ever run.

After two years on Supabase, I moved all my new projects to Neon and wrote down why: they are an agents-first company that ships everything to the API, they develop at very high velocity, and branching.