Software-agent research is getting more executable, replayable, and production-bound

The strongest work today tries to grade coding agents on signals that can survive real engineering conditions. ProdCodeBench builds tasks from actual developer-assistant sessions in an industrial monorepo, keeps the original prompts, backs out the landed diff, and scores agents with stable fail-to-pass and pass-to-pass tests. That gives offline evaluation more of the constraints teams face in practice. ClickHouse adds a field report from the deployment side: agents are useful when they can read code, run tools, and stay inside a review-and-test loop. The numbers there are operational, not benchmark-grade, but they show why production teams care about grounded evaluation at all.

Several papers focus on what happens after a model makes a plausible first move. Beyond Resolution Rates finds that many failures are not search failures in the narrow sense. Agents often locate the right file and still miss the fix because architectural judgment is wrong. ToolMisuseBench looks at a different layer of reliability: tool calls under schema drift, timeouts, rate limits, and authorization errors. Its baselines recover on some timeout and schema faults, but overall success stays at 0.250 and budgeted success is flat. The common message is that agent quality now depends on validation behavior, recovery policy, and action discipline, not just answer quality.

Testing is becoming a first-class generation target, not just a post hoc check. TestDecision treats test-suite construction as a sequential optimization problem and reports large gains in branch coverage, execution pass rate, and bug finding over base open models. APITestGenie applies the same execution-minded attitude to APIs: it starts from business requirements plus OpenAPI specs, generates runnable Jest and Axios tests, and reaches at least one valid script for 22 of 25 requirements. Together these papers make testing more executable, more requirement-aware, and more useful as a direct product of the model pipeline.