Robert Ta introduces Vercel's Eve, an open-source framework for AI agents that handles durable execution, sandboxing, and human approvals out of the box. Over 100 production agents already run on it. The framework lets developers focus on what an agent does instead of building infrastructure from scratch.
Vercel's take on Next.js for agents is called Eve. In the launch, they showed more than 100 production agents already running on it. It's an open-source framework for building, running, and scaling agents. Here's the idea: building an agent should mean defining what it does without having to assemble every piece it needs to run in production. So Eve comes with that production layer built in. Durable execution, sandboxed compute, human-in-the-loop approvals, subagents, evals, all of that stuff you would otherwise have to put together. And the core design principle is an agent is a directory. You've got agent.ts, the model it runs on, instructions.md, who it is, tools, what it can do, skills, what it knows, subagents, who it delegates to, channels, where it lives, and schedules, when it acts on its own. It's a pretty clear way to organize the pieces. I like this idea of a Vercel-like experience for agents. You need somewhere to ship them and manage them. And as more of these harness frameworks come out, look at the choices they've made and ask how they fit the thing you're building. What should your harness actually be responsible for? One thing worth looking at is that Vercel had already been using it internally and with clients, so there's experience behind those patterns. The file structure itself isn't magic. It's a pattern you can understand, use, and build on. That gives you a starting point for the parts you don't need to figure out from scratch. Then the interesting question is where you want to put your own effort. Which parts of your stack are the things you really want to innovate on? You can take those existing patterns and use them as a basis for thinking through your architecture instead of spending all your time rebuilding the same pieces. For a harness like ours, that brings it back to human domain expertise. How do we capture that? How do we use it in evals? How do we build an LLM judge and keep improving the agent inside that loop? That's the work that takes you toward a system that can actually add value for customers.
