Coding-agent research now treats verification surfaces as core system design

Across the week, strong papers kept pulling agent behavior into forms that can be checked outside the model. Requirements, contracts, architecture descriptors, and typed failure signals all act as hard surfaces for planning and review. The practical gain is clearer repo-scale work and fewer hidden assumptions in the loop. ReCodeAgent and REAgent tie generation to explicit task descriptions, while later work extends the same idea to formal specs and durable write paths.

Evaluation tightened around exact edits, end-to-end tasks, and runtime evidence. That changes what counts as success. Papers this week ask whether an agent can complete a repository task, follow design rules, expose what evidence it used, and stay within cost or action limits. Benchmarks such as CLI-Tool-Bench, SWD-Bench, and HiL-Bench point in the same direction: useful scores come from long tasks with visible constraints, not small isolated completions.

Verification now sits inside the action loop for both coding and security work. Tests, runtime instrumentation, execution grounding, and guardrails are used as gates before an agent commit or repair is accepted. This week also kept a realistic view of limits: security work reports exploitable generated code and weak reliability in messy multi-step settings, even as systems like DeepGuard and Verify Before You Fix improve the checking path.