Repository-grade coding exposes where agents still break
The day’s clearest signal is stricter evaluation for large language model (LLM) software agents. Generated code must satisfy architecture, tests, migrations, and real developer traces. TACT and ASTOR improve control.