Software-agent research is converging on verifiable training loops and hard control gates

Reinforcement learning is getting more usable because the training target is narrower and the environment is faster. Scaling Coding Agents via Atomic Skills breaks software work into five verifiable skills and reports an average 18.7% gain across ten tasks, including SWE-bench Verified at 0.585 from 0.507. SandMLE applies the same execution-first logic to machine learning engineering: synthetic tasks cut runtime from about 200 seconds to under 15 seconds, then improve Any Medal rate by 20.3% to 66.9% on MLE-bench-lite. The common pattern is simple: make each step cheap to run and easy to score, then scale on-policy RL.

Repository agents are being judged on whether they can repair behavior, not just patch files. Agent-CoEvo searches over code patches and test patches together, then scores them through a pass-fail matrix. It reports 41.33% resolved on SWE-bench Lite and 46.4% on SWT-bench Lite, with higher test quality than the listed baselines. In microservices, Mirage keeps the model in the dependency loop during testing and reaches 99% status-code fidelity and 99% response-shape fidelity across 110 scenarios. Both papers put the model inside an executable feedback loop where behavior is checked as it unfolds.

A second line of work is reducing runtime model freedom once the task is clear enough to compile or verify. Compiled AI generates a small code artifact once, validates it, and then runs deterministically. On BFCL it reaches 96% task completion, breaks even after about 17 transactions, and runs at 4.5 ms median latency with 100% reproducibility. Nidus pushes harder on governance: every change is checked against a solver-backed living specification, and the paper reports a self-hosted 100,000-line system with proof obligations on every commit. This day’s strongest systems do not ask the model to be careful on its own. They bind it to code, tests, or formal checks.