OpenAI published a September 2026 report describing how coding agents are being used inside its research organisation. The report focuses on research workflows rather than presenting coding agents as a replacement for researchers.
What changed in the workflow
OpenAI says researchers are writing more code and running more experiments with the help of coding agents. It also describes a shift in the kinds of tasks agents are used for, as researchers become more comfortable delegating parts of technical workflows.
That matters because AI-assisted programming is increasingly about more than autocomplete. Coding agents can help with multi-step tasks such as exploring a codebase, writing changes, running experiments and iterating on results.
What developers should take from it
The useful takeaway for students is not simply “use an AI coding agent.” Instead, learn a workflow where you remain responsible for the result:
- Give the agent a small, well-defined task.
- Ask it to explain assumptions before making risky changes.
- Review the generated code.
- Run tests and inspect failures.
- Keep sensitive credentials and production access out of the agent's reach.
Why this trend matters
Coding agents are becoming part of the software-development toolchain. That makes fundamentals such as Git, debugging, testing, system design and code review even more useful because they give developers the ability to verify what an agent produces.
Source
Published 6 September 2026. This post summarises OpenAI's own report and its stated observations about internal research workflows.