I recently interviewed Peter Gostev, AI Capability Lead at Arena.ai. He created what I call a poop test, officially known as the BullshitBench. But something else he mentioned caught my attention.
This is a notable example of a real-life AI agent. While editing photos in Adobe Lightroom, I found 50 images that needed denoising. This is usually a tedious task that is done one at a time. Instead of doing it manually or knowing how to do batch processing, Gostev turned to OpenAI’s Codex AI coding service to figure it out.
“If you want to remove noise from 50 photos, you have to click them one by one. That sounds like a lot of work, so we asked Codex to come up with a way to do it,” Gostev told me. “It worked.”
What’s remarkable is how Codex achieved this. Rather than using official APIs, plugins, or browser workarounds, this was achieved by somehow connecting directly to the desktop app, despite the lack of clear support for it.
Gostev is technically advanced and what he did here is probably something I can’t do. Still, this is a glimpse of where AI agents are headed. AI agents don’t just help, they autonomously navigate and interact with software just like humans (in this case, faster and better than smart humans).
Sign up for BI’s Tech Memo newsletter here. Please contact us by email. abarr@businessinsider.com.