OpenAI has published research on how coding agents, AI systems that write and run code with limited human direction, are contributing to the pace of AI research itself. The work examines where these tools speed up research tasks and where they fall short. The results are uneven: agents perform well on some tasks and poorly on others, a pattern often described as a 'jagged frontier'. The findings matter because they bear on recursive self-improvement, the idea that AI systems could help build more capable successors, creating a self-reinforcing loop of faster progress. Uneven performance suggests that any such acceleration would be patchy rather than smooth, and that humans would still need to judge which tasks agents can handle reliably. Whether that loop is close remains an open question, and gains that arrive unevenly make it harder to predict. For most Australian businesses and professionals, this research changes nothing immediately. It is a look at frontier lab research, not a new product, price change or regulatory requirement. Teams already using coding assistants may find it a useful reminder to check agent output carefully on tasks where results are inconsistent, since errors in code can be costly. If AI-generated code or agent workflows handle personal information, obligations under the Privacy Act 1988 still apply.