OpenAI published a field report on July 28 documenting eight real-world projects where research teams used Codex and, in several cases, Claude Code alongside it, to modernize scientific software that had been quietly rotting for years — the kind of code written by a couple of grad students, packaged carelessly, and then relied on by an entire subfield without ever getting proper engineering investment. The results were striking in places: a from-scratch Rust rewrite of a 20,000-line C/C++ genome aligner that matched the original's output 99.8% of the time, a GPU-native redesign that cut a synthetic genome generation run from 1,610 seconds to 27, and a 60x speedup in RNA-sequencing quality control. But the report's real value for developers isn't the speedup numbers, it's the workflow pattern the contributors converged on independently across eight unrelated projects. Every team split the work the same way: let the agent handle the well-scoped, mechanically verifiable parts (dependency upgrades, build system replacement, direct language ports) while a human stayed responsible for defining what "correct" means and building the harness to check it, usually via exact-output parity against the old implementation, agreement with an external reference tool, or pre-computed answers on synthetic data. Agents consistently produced confident-sounding output even when it was wrong, so the projects that succeeded were the ones that had an automatic, objective pass/fail check rather than relying on the agent's own self-assessment. Teams also converged on iterating in small, benchmarked stages instead of attempting one giant rewrite, because the "last mile" of matching subtle numerical edge cases ate up far more time than the initial implementation. If you're using agentic coding tools for anything beyond toy projects, this maintain-verify split — agent does implementation, human owns the acceptance test — is the pattern worth adopting deliberately rather than discovering the hard way.