Field notes, not thought leadership.

Why most AI agents fail in production (and the boring fixes that work)

The failures are rarely about model quality. They're about the unglamorous parts nobody wanted to own.