Vercel shipped an agent framework on June 17, and the first thing you notice is that it doesn't look like one. There's no orchestration graph to declare, no state machine to wire. An agent in eve is a directory of files: an instructions.md for what it should do, a tools/ folder of TypeScript functions, a skills/ folder of more markdown, and a subagents/ folder of the same shape one level down. Vercel's pitch is a single analogy — Next.js for agents — and the analogy is precise. eve is to the AI SDK what Next.js is to React: the SDK gives you the primitives, eve gives you the conventions and the production plumbing already bolted together.
The temptation is to line eve up against LangGraph and score the loop. Don't. That comparison is already over.
The loop converged; the harness is the whole fight#
Both of these run a tool-calling agent. Both checkpoint state so a session survives a crash. Both pause for human approval and resume. LangGraph models this as a directed graph of nodes and edges over typed state; eve models it as a directory that compiles down to a durable service. Different surfaces, same machine underneath. If you choose between them by asking which one can express my control flow, you will flip a coin, because the answer is "both."
The decision lives one layer out — in the part nobody puts on the feature table because it isn't a feature, it's an entire stack. Durable execution. A sandbox to run model-written code without it touching your box. A broker that holds credentials so your agent can call a third-party API without you pasting a key into a prompt. Tracing you can read. Evals you can run in CI. Somewhere to actually deploy the thing. This is the harness — the reliability layer wrapped around an unreliable model — and assembling it is where agent projects quietly go to die.
The loop is commodity. The harness is the product. eve's only real innovation is admitting that out loud.
LangGraph's answer to the harness is honest and total: it doesn't ship one. It hands you a graph object, provider-agnostic, that runs wherever Python runs, and the rest is yours. Persistence backend, retry policy, the secrets story, the sandbox, the place it lives — you assemble all of it. That's not a gap; it's the deal. LangGraph is a library, and a library's virtue is that it makes no claim on the world around it.
eve makes the opposite trade. It ships the harness as defaults — six production capabilities turned on the moment you scaffold an agent: durable execution, a sandbox, human-in-the-loop approvals, subagents, OpenTelemetry tracing, and a built-in eval runner. You don't choose them. You inherit them. eve is a harness sold as a framework, and that is genuinely new — not because the pieces are new (durable workflows, sandboxes, and gateways have existed for years) but because someone finally pre-wired them into the unit of work and called the bundle the product.
The directory is a runtime contract in disguise#
Here is the part the analogy hides. A convention — "an agent is a directory of files" — is only worth anything if a runtime enforces it. Next.js's file-based routing means nothing without the Next.js build and server turning app/page.tsx into a route. eve's directory means nothing without the runtime that turns instructions.md and tools/ into a durable, sandboxed, observable service.
And that runtime is Vercel's. eve orchestrates components Vercel already builds — the AI Gateway for model routing and fallback, the Workflow SDK for durable execution, the Sandbox for isolated compute, Vercel Connect for brokered credentials. The durability you get isn't generic: a session is an event log that deterministically replays to reconstruct state, so the agent can park while it waits on a message or a tool result and resume after a cold start or a redeploy. That is excellent engineering. It is also the most specific kind of lock-in there is, because it's invisible — it lives in defaults you never opted into.
You can run eve locally. It supports Docker, a microsandbox, or plain bash as sandbox backends with no Vercel account at all. But the production stack — Workflow, Sandbox, AI Gateway, the Agent Runs dashboard — is proprietary. eve is, by some distance, the only major agent framework you cannot fully self-host, and that's not an oversight Vercel will patch. It's the business model. The harness is the value; the platform is how the value gets billed. If portability is the requirement, Mastra is the TypeScript framework that ships durable execution (via Inngest) and still runs on your own hardware — and LangGraph is the answer when the agent simply must not belong to one vendor.
So pick the thing you're actually choosing#
State the decision as what it is, and it answers itself.
- You want the harness handed to you. You're a TypeScript team, you're on Vercel or glad to be, and you would rather inherit durability, sandboxing, auth brokering, tracing, and evals than spend a quarter assembling them. You accept that "deploy" and "Vercel" become the same word. That's eve. The agent is a directory of files; the floor it stands on is rented, and the rent buys you speed.
- You want to own the stack. Your agent runs on-prem, or across clouds, or somewhere no hyperscaler reaches. You're Python-first, or you already have an observability pipeline and a secrets vault and an opinion about sandboxes. You'll pay in assembly for the right to move. That's LangGraph. The graph is yours, and so is every wire you run to it.
Both are correct. What's incorrect is benchmarking them on the agent loop, because the loop is the one thing they agree on. The frameworks converged on how an agent thinks so the platforms underneath could diverge on who owns it. eve's wager is that most teams will trade the keys for the kitchen — that a harness you don't have to build is worth a runtime you can't take with you. For a lot of production agents, that wager is going to pay. Just sign the lease knowing it's a lease.
