> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # OpenAI says coding agents reshape its own AI research - URL: https://www.notatechguy.com/openai-says-coding-agents-reshape-its-own-ai-research/ - Published: 2026-09-06T18:18:42.000Z - Updated: 2026-09-06T18:18:42.000Z - Description: OpenAI says coding agents are reshaping its internal AI research, with early data on agent usage and experiment velocity, but no verifiable metrics. - Author: Marcello Babbili - Tags: Technology & AI, OpenAI On September 6, OpenAI published "Research acceleration: The view inside OpenAI," a first-person account claiming that coding agents are reshaping how the company conducts its own AI research [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). The post promises early data on how often agents are used, how fast experiments move, and how complex the delegated tasks are [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). What the title alone cannot tell you is whether any of those numbers can be checked by anyone outside the building. **My read:** This is the third OpenAI publication in under a year that frames AI as a scientific accelerator, and it is the most self-referential yet. The November 2025 piece on GPT-5 science experiments [P³](https://openai.com/index/accelerating-science-gpt-5/?ref=notatechguy.com) and the January 2026 "AI as a Scientific Collaborator" report [P⁶](https://cdn.openai.com/pdf/f4b4a5da-b2de-418d-9fcd-6b293e9dc157/oai%5Fai-as-a-scientific-collaborator%5Fjan-2026.pdf?ref=notatechguy.com) at least pointed outward, to external scientists and ChatGPT users. This one points inward, at OpenAI's own researchers using OpenAI's own tools. That makes it harder to separate the finding from the sales pitch. I want to see the actual data before I accept the claim that research is being "reshaped." ## The pattern behind the post The September 6 article sits in a sequence of OpenAI publications making the case that AI accelerates science. In November 2025, OpenAI published "Early experiments in accelerating science with GPT-5," describing insights from partnerships with scientists in areas ranging from health to energy [P³](https://openai.com/index/accelerating-science-gpt-5/?ref=notatechguy.com). In January 2026, it released "AI as a Scientific Collaborator," noting an average of 8.4 million weekly messages on advanced hard-science and math topics, with about 1.3 million weekly ChatGPT users concentrating on advanced math and science [P⁶](https://cdn.openai.com/pdf/f4b4a5da-b2de-418d-9fcd-6b293e9dc157/oai%5Fai-as-a-scientific-collaborator%5Fjan-2026.pdf?ref=notatechguy.com). That report said monthly advanced science messages grew nearly 50% [P⁶](https://cdn.openai.com/pdf/f4b4a5da-b2de-418d-9fcd-6b293e9dc157/oai%5Fai-as-a-scientific-collaborator%5Fjan-2026.pdf?ref=notatechguy.com). Separately, OpenAI has been publishing the mathematical results its internal models produce. A paper titled "Ten Advances in Mathematics and Theoretical Computer Science" presents results from an internal model, including a precise calculation of the asymptotic strength of the Cohn-Elkies linear program applied to high-dimensional sphere packing [P⁵](https://cdn.openai.com/pdf/ten-proofs-oai.pdf?ref=notatechguy.com). That is a real mathematical result, attributed to an internal model, in a field where verification is possible because other mathematicians can check the proof. The new post is different. It is not about external scientists using ChatGPT, or about a model solving a specific problem. It is about OpenAI's own researchers using coding agents to do AI research faster [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). The claim is qualitative: coding agents are "reshaping" research [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). The data is described as "early" [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). ## What "coding agents" means here Coding agents are AI systems that write and run code with limited human oversight, then debug what breaks. Inside OpenAI, the claim is that these agents are changing how experiments are designed and executed [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). The post covers agent usage, experiment velocity, and task complexity [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com), which suggests OpenAI is tracking how often researchers delegate coding work to agents, how fast experiments move when they do, and how hard the delegated tasks actually are. This connects to OpenAI's broader tooling push. On the same day as the research post, OpenAI's Codex CLI, a command-line coding agent, was trending on GitHub with 121,803 stars. The company also maintains gpt-oss, two open-weight models (gpt-oss-120b and gpt-oss-20b) on GitHub with 20,363 stars and an Apache 2.0 licence [P⁴](https://github.com/openai/gpt-oss?tab=readme-ov-file&ref=notatechguy.com). The tools researchers use internally and the tools OpenAI ships externally are likely related, though the post does not confirm which specific agents it refers to. ## What to do about it If you run a research team or an engineering org, the claim worth testing is not "AI makes us faster" but "coding agents change which tasks we attempt." A genomics lab that currently scripts its own pipeline analysis by hand could pilot an agent that writes and runs experiment code overnight, then compare the number of experiments completed per week with and without the agent. The measure that matters is not lines of code generated but experiments finished and validated. The practical step: pick one repetitive coding task your team does by hand this week, hand it to a coding agent, and time the result. If the agent completes it correctly in under an hour, you have your own early data point. ## What we don't know yet The evidence for this story is thin by design. The full article text was not reviewed; only the title and description were available [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). OpenAI has not published specific quantitative metrics proving research acceleration, and no independent researcher has verified that coding agents are reshaping AI research at OpenAI. The claim that research is being "reshaped" is qualitative and self-reported [S¹](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com). The broader question is whether OpenAI's self-reported agent usage data will be backed by verifiable outputs, the way the sphere-packing result in the Ten Advances paper can be checked by mathematicians [P⁵](https://cdn.openai.com/pdf/ten-proofs-oai.pdf?ref=notatechguy.com). Until the full article is reviewed and its data assessed, the post is a signal about OpenAI's direction, not a proven finding. The next signal: OpenAI's next research publication. Based on the cadence from November 2025 through September 2026, that could come before year-end. We will check whether it includes agent-usage data that holds up outside OpenAI. If you want that follow-up in your inbox, subscribe now and you will not miss it. --- *Sources: [S1 — Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai?ref=notatechguy.com) · [S2 — Research acceleration: The view inside OpenAI - OpenAI](https://news.google.com/rss/articles/CBMidEFVX3lxTE1PaU1kVmVkamQtbXI0aE5aU2plRHlhT0dsbi1VTDlJTFBsNEZOQUVpdENyWVB3ZTZ6TFFSakp3RGFmbXprVl94QnZVMVR6OUEzY1c1ZG1YdF9naWZEb0ZqbHQ0Y1c2dnljWjhZeDVmbC1TTEdQ?oc=5&ref=notatechguy.com) · [P3 — Early experiments in accelerating science with GPT-5 | OpenAI](https://openai.com/index/accelerating-science-gpt-5/?ref=notatechguy.com) · [P4 — openai/gpt-oss](https://github.com/openai/gpt-oss?tab=readme-ov-file&ref=notatechguy.com) · [P5 — Ten Advances in Mathematics and Theoretical Computer Science](https://cdn.openai.com/pdf/ten-proofs-oai.pdf?ref=notatechguy.com) · [P6 — AI as a Scientific Collaborator January 2026](https://cdn.openai.com/pdf/f4b4a5da-b2de-418d-9fcd-6b293e9dc157/oai%5Fai-as-a-scientific-collaborator%5Fjan-2026.pdf?ref=notatechguy.com)* ## More from Not A Tech Guy - [OpenAI Codex CLI trends on GitHub with 121,803 stars](https://www.notatechguy.com/openai-codex-cli-trends-on-github-with-121-803-stars/) - [ACToR retrieval lifts AI code generation 15% on CoderEval](https://www.notatechguy.com/actor-retrieval-lifts-ai-code-generation-15-on-codereval/) - [AI agents learn your standards, up to 20.9% better](https://www.notatechguy.com/ai-agents-learn-your-standards-up-to-20-9-better/) --- *Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.* ![OpenAI's AI-for-science publication timeline](https://storage.ghost.io/c/6e/89/6e896869-22ef-4281-a213-b4c462c17cff/content/images/2026/09/chart_c0ac188bbae2ed0c31cc.png)