Eric Siu built Picasso, an AI agent that generates advertising creatives at scale while respecting brand guidelines and competitor analysis. The team prototyped an outreach flow using OpenAI Codex to create on-brand ad examples, but hit reliability issues when resource limits caused the agents to fail during the build.
So my team actually made a design, a creative agent called Picasso that will spin up creatives at scale. Okay. Okay. And like a lot of people can do that, but ours is focused on, hey, what's actually performing? What are your brand guidelines? And also what are your competitors doing? And we have a designer in the loop. And I was like, huh? So I took it for myself. I dumped it into Codex, which is OpenAI's, and I said, hey, I want you to basically make this into where I can send Neil cold emails or Noah cold emails and say, hey, Salesforce, we made some ads for you over here that are on brand. Here's 3. Would you like more? Right. And so that I built with Codex. And then it told me to finish off the build using Hermes. OpenCLaw and Hermes fail because if you add too much to its disk or too much to its memory, it starts to— I can't find it anymore, Neil. I don't know. I don't have that capability. So you start to go back and forth with it. It starts to become really unreliable.
