Codex used Lightroom UI
Business Insider reports OpenAI’s Codex was able to operate Adobe Lightroom like a human to denoise photos without a plugin or formal API, using the application UI directly to complete tasks (businessinsider.com). The example highlights agents automating UI-driven workflows where no official API exists (businessinsider.com).
A coding agent usually works through application programming interfaces, the software hooks developers expose for automation. In this case, OpenAI’s Codex reportedly handled Adobe Lightroom by using the app’s own buttons and menus to denoise 50 photos. (dnyuz.com) Business Insider’s account, republished by DNYUZ on April 11, said Peter Gostev, artificial intelligence capability lead at Arena.ai, gave Codex a Lightroom job he did not want to do by hand. Gostev said the task was clicking through 50 images one by one to run denoise. (dnyuz.com) The reported detail that stands out is what Codex did not use. The story said there was no official application programming interface, no plugin, and no browser workaround involved in the Lightroom run. (dnyuz.com) Adobe’s denoise tool is itself an artificial intelligence feature built for noisy raw photos, especially high-ISO images shot in low light. Adobe introduced Denoise in April 2023 for Camera Raw 15.3, Lightroom 6.3, and Lightroom Classic 12.3. (blog.adobe.com) OpenAI describes Codex as a coding agent that can write code, explain codebases, review code, debug failures, and automate repetitive development work. OpenAI also says Codex can connect to outside tools and documentation through Model Context Protocol servers in its command-line interface and integrated development environment extension. (developers.openai.com, developers.openai.com) That makes the Lightroom example less about photo editing than about software control. If an agent can inspect a desktop interface, infer the right sequence, and execute it, then tasks inside programs without a formal automation layer become candidates for the same treatment. (dnyuz.com, developers.openai.com) The limits are also visible in the reporting. Gostev was the person running the experiment, and the public description does not spell out the exact setup Codex used to see Lightroom, send clicks, or verify each exported result. (dnyuz.com) Still, the example points to a simple shift in where automation can happen. Instead of waiting for a plugin or an application programming interface, the agent reportedly used Lightroom the same way a person does, only at batch scale. (dnyuz.com)
