Coding-agent research is getting judged by execution loops, runtime evidence, and real action control
This day’s strongest papers make coding-agent claims more executable and more inspectable. The emphasis is concrete control over what agents read, remember, run, and are allowed to change.