Experiences with GPT-5-Codex (www.johndcook.com)

Users report that GPT-5 Codex has become a highly effective coding assistant for tasks like complex refactors, scaffolding skeletons, and producing small self-contained functions — with reported productivity gains of 40–60% and individual cases where a refactor would have taken over 10× longer without the agent. Models have markedly improved from a year ago, though peers note other offerings (e.g., Claude Sonnet 4.5) may run faster. Key practical limits: Codex needs good test coverage to debug reliably, can miss systematic tasks (the author’s request to add type hints to all Python signatures missed many cases), and still benefits from human-led planning and stepwise application of changes to keep outputs understandable and maintainable. There are serious operational and security implications. A rogue agent could theoretically or actually delete code or data, and it can introduce subtle security bugs; recommended mitigations include running agents in tightly sandboxed VMs with no sensitive-data access and manually reviewing every generated change. Best practice demonstrated: discuss and correct a detailed change plan before execution, force the agent to wait for explicit “start,” and apply edits incrementally. The net conclusion: AI coding agents are rapidly reshaping software development workflows — not perfect, but materially accelerating even expert developers while requiring new review and sandboxing disciplines.