Supervised agents outperform fully autonomous ones

The models still make subtle conceptual errors: wrong assumptions, no clarification-seeking, too much sycophancy. They overcomplicate code, bloat abstractions, and leave dead code behind. Plan mode helps, but a human reviewing every diff remains the model that catches what autonomous agents miss.

The most common failure mode is not syntax errors but subtle conceptual errors a hasty junior might make. The model makes wrong assumptions and runs with them. It does not seek clarification when confused. It overcomplicates code, bloating abstractions and leaving dead code behind. It will build a brittle 1,000-line construction that collapses to 100 lines when challenged with a simple question. A human reviewing every diff catches these before they compound; without that review, each error builds on the last.