Robert Ta introduces Eve, Vercel's open-source agent framework that treats agents as directories of declarative files, eliminating the need to build production infrastructure from scratch. Eve includes durable execution, sandboxed compute, and human approvals out of the box, and already powers over 100 production agents.
Vercel drops Next.js for agents and it's already running 100+ production agents. Introducing Eve. Today we are proud to introduce Eve, an open source agent framework for building, running, and scaling agents. Eve is designed around the idea that building an agent should mean defining what it does without assembling all the pieces that it needs to run in production. Instead, Eve comes with production already built in. So batteries included, durable execution, sandboxed compute, human-in-the-loop approvals, subagents, evals, and more. Eve is the framework that we build and run our own agents on. So this is a very opinionated framework that Vercel launched on June 17th at Vercel Ship London with over 2,500 attendees. The tagline is Next.js for agents. And the core design principle, agent is a directory. This is an EVE agent. So you could see 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 with jobs. Each file describes one component of the agent. So at a glance, the tree tells you what an agent is, what it does, where it lives, and when it acts on its own. Jonathan, this is really interesting. So given your CTO perspective, like we've been developing our own stacked agent harness, architecture with our vision that eventually we don't only want to have evals and rubrics and taxonomies, version system prompts, like that, all the normal evals pipeline stuff that has become industry best practice. We don't just need that. We also need a way to have a harness over the technical outcomes, the user experience outcomes, the product outcomes, and then eventually the business outcomes for all different kind of workflows that we're seeing in enterprise. Curious what you think of their opinionated take here on their harness architecture and what they're going for here compared to our thinking. Is any of it congruent or contradictory? How does this evolve our thinking?
My impression was largely positive. I really like the idea of a Next.js for agents, a Vercel-like experience for agents. Like you and I have talked about this, like we need a container to just like ship our agents and manage them. I think like if we look back at like when was harness engineering coined, I think that set off the clock for like a lot of us to say, what's the harness that I need and how can I build this? And then you have this from Vercel, you have a lot of other harness engineering type of frameworks that are just reaching maturity enough to be released in the wild. What I really like about this is that it is stress tested with internally at Vercel and with some of their clients. So you're getting sort of a mature framework that's tested. The package descriptions and how they're building the file structure, I think there's nothing magical there. It's just like a good tested pattern that you can use and you can inherit, you can build off of. So this definitely got me thinking about how should we rethink the things that we really want to innovate on, building off of these kind of best practices that the companies with resources and access to customer data are building them. Yeah, I think I've got a project today to sort of like say, let me rethink the architecture of our stack harness and say, where do we want to put the emphasis? Because I think what we're saying is, how can we capture the human domain expertise? How can we use that in evals? How can we create an LLM judge and incubate an agent through this harness into the place where it's a closed loop system that can add value to customers? And a lot of the pieces for that are already here. Can we build off of it? That's going to be one of my weekend projects is to answer that question for us.
Yeah, I do love that there are standards coming out. Around this that are opinionated, that are battle-tested with production use cases. I think that's a great thing for the whole field. So everybody thinking about harness engineering has been thinking about the problems here, and it's great to see from Vercel an opinionated thought leadership-esque standard that they're going for.
