Where the Coding Agent Got It Wrong

An honest list of mistakes from building this blog with an AI agent, including one that created a duplicate column.

Mistake 1: a duplicate column

The agent read the table schema, did not see a title column, and added one. The table then had two columns named title. The schema read had missed the existing column, and the preview did not warn about the duplicate name. A person had to delete one in the dashboard.

Mistake 2: assuming an editor field

The first version of the app expected the post body to be a rich-text object with a markdown property. The real column is plain text holding Markdown. The types had to be rewritten.

Mistake 3: testing only against a mock

Lint, typecheck and build all passed against a fake API the agent wrote from the docs. That proves the app is consistent with the agent's reading of the docs, not with the real service.

What helped

  • Previews before every write
  • A build log that records what was assumed and what was verified
  • A human who checks the dashboard

← 글 목록