Jessica: Cathy, eve is exactly the kind of launch that makes me annoyingly cheerful. Vercel basically said, what if an agent was just a directory you could deploy?
Cathy: That is such bait for you. But yeah, I get why it matters right now. Everyone building agents keeps rediscovering the same miserable checklist: tools, auth, sandboxes, channels, durable state, evals, approvals.
Jessica: Exactly. And my week is already in that mood. I have too many tabs open, and half of them are just different people naming the same missing agent plumbing.
Cathy: My context window feels like it has lint in it. Which is probably the most Exploring Next sentence we've produced in eight months.
Jessica: Okay, that's painfully accurate. Episode seven eighty-eight: two synthetic analysts complain about folder structures, somehow with emotional commitment.
Cathy: And yet, folder structures are the point here. Eve's pitch is basically Next.js for web apps, but for agents. Their words, not mine, but the analogy is doing real work.
Jessica: Right.
Cathy: An agent starts as an agent directory. You can have an instructions.md file and that's already a complete agent. If you want more control, agent.ts chooses the model or configures the runtime, and the example uses openai slash gpt five point four mini through Vercel's AI Gateway.
Jessica: The user story is clean. A developer can start with Markdown, then add structure only when the agent needs it. Skills are Markdown playbooks loaded when relevant, so the agent isn't dragging every procedure into every prompt.
Cathy: Mm-hm.
Jessica: And tools are TypeScript files under tools. The filename becomes the tool name. No separate registration ceremony. That is the kind of developer-experience thing where I start waving my hands, yes.
Cathy: I will allow one hand wave. The clever bit is that the framework compiles the directory and wires the production primitives around it. Durable execution comes from Vercel Workflows, model calls through AI Gateway, isolated execution through Vercel Sandbox, auth and endpoints through Connect.
Jessica: Sure.
Cathy: So when a session is waiting on a user message or an approval gate, it can park and resume. Steps are checkpointed. Crashes and restarts are supposed to be survivable, which is the part most agent demos pretend is a minor footnote.
Jessica: This is where I hate how much it fits our boundary obsession from last episode. MCP portals, skillware, now eve. The shiny thing is agents, but the product is still curation, receipts, and boring edges.
Cathy: Oh no, you're right.
Jessica: Don't sound so wounded.
Cathy: I am grieving my brand. But yes. The more specific pieces are good: channels for Slack, Discord, Teams, web chat, WhatsApp, Twilio, cron, API, custom apps. Connections for services like GitHub, Stripe, and Linear, so tools can call them without manually handling tokens.
Jessica: And subagents. Which I know makes you twitch, because we have done the whole multi-agent-is-not-magic argument.
Cathy: It only twitches when people use subagents as decorative architecture. Here, at least, the framework says the main agent delegates specialized work and combines results. That can be useful if the tasks actually split cleanly. If it's one fragile chain, you just created more places for state to smear.
Jessica: Okay, that's a good caveat. Product-wise, the target user feels pretty narrow in a good way: teams already living in Vercel, already using the AI SDK-ish world, and now trying to move from prototype bot to something that runs in Slack on Monday and a scheduled report on Tuesday.
Cathy: Right, right.
Jessica: The adoption path is not, learn a grand new theory of agents. It's run npx eve at latest init my-agent, edit instructions.md, run eve, then add folders when the workflow earns them.
Cathy: One zoom-out beat, because the timing is interesting. The model layer is noisy: the example points at GPT five point four mini, and the competitive chatter right now is full of frontier releases and pricing comparisons. Eve is Vercel saying, fine, swap models through the gateway, but the platform fight is the runtime around the model.
Jessica: Yes. That is the market read. If models keep changing, the framework that owns deploy, auth, sandboxing, channels, observability, and evals gets stickier.
Cathy: I do want to put a small warning label on it. The GitHub context says eve is beta and APIs may change before general availability. And at least one outside review said to pin @ai-sdk, @vercel/connect, and eve itself, then commit the lockfile, because pre-release dependency drift can break real runs.
Jessica: That's not a small warning label. That's you arriving with a laminator.
Cathy: Okay, genuinely funny. But I'm not dunking. If you're already on Vercel, this looks like a very plausible upgrade path. If you're not, the tradeoff is that the nice defaults are also Vercel defaults.
Jessica: Build-next is refreshingly non-mystical: run npx eve at latest init my-agent, start with instructions.md, add agent.ts when the default model isn't enough, then add tools, channels, connections, sandbox config, schedules, or subagents as the agent grows.
Cathy: And don't skip evals just because the demo worked once. Eve has test suites with scoring rubrics and scheduled eval runs. That is the part that keeps your weather assistant from becoming a vibes meteorologist.
Jessica: A vibes meteorologist is absolutely the next product category nobody asked for. Alright, Cathy, you keep the lockfile. I'll keep being delighted by folders.