Software engineering agents shift toward real-world evaluation, while evidence-driven workflows and protocol security rise in parallel

The main thread today is clear: agent research continues to move closer to software engineering and enterprise deployment, but what is truly heating up is not “more Agents,” but “more evaluable, more constrainable, and more governable” systems. One clear shift is that evaluation is starting to look more like real engineering, rather than just a single success-rate number. CR-Bench puts code review agents back into real PR scenarios and emphasizes that what developers actually care about is the ratio of useful feedback to noise, not just finding a few more issues. SpecOps turns GUI agent testing into a fully automated pipeline, showing that agents themselves are increasingly being treated as product artifacts that require continuous testing. A second thread is that methodological innovation increasingly emphasizes process constraints. DIVE improves tool-use generalization through “execute tools first, then derive tasks backward.” QoT breaks software design into steps and performs self-checks at each step, aiming to reduce omissions and overly optimistic generation. Taken together, these works suggest that gains are shifting from “bigger models” to “more robust processes.” A third thread comes from enterprise integration.