Code agents enter real engineering loops: repository understanding, end-to-end evaluation, and safety governance heat up

This week’s software engineering and code intelligence research has a very clear main thread: code agents are shifting from “can generate” to “can execute, verify, and operate over time in real repositories.” The true competitive frontier has become repository understanding, end-to-end evaluation, memory management, and safety governance. One obvious change is that research is discussing less and less whether a single generation looks good, and more and more whether agents can complete a closed loop in real engineering settings. RAIM targets repository-level feature addition. BeyondSWE expands tasks to cross-repository work and dependency migration. Echo connects retrieval, execution, and verification. This shows that code agents are starting to be designed around development workflows rather than single-problem benchmarks. The second signal is that evaluation is getting tougher. VibeCodeBench requires delivery of complete web applications. SWE-CI focuses on continuous maintenance. The materials repeatedly point out that once real integrations like payments, email, and databases are involved, model performance drops sharply. Evaluation no longer asks only “was something written,” but “can it be deployed, can it keep evolving, and did anything break after the change.” The third signal is that engineering infrastructure is heating up.