AI Agents

Vercel's eve Bets That an Agent Is Just a Directory

Published Jun 26, 2026AI Agents

Vercel released eve, an open-source framework that treats an AI agent as a directory of files and bundles durability, sandboxing, and approvals as conventions. It's a "Next.js for agents" bet — with a vendor-lock-in caveat.

A new framework with an old idea

On June 17, 2026, at its Ship 2026 conference in London, Vercel released eve, an open-source framework for building production AI agents [1]. The pitch is unusually concrete: an agent is a directory. You define one as a folder of files on disk, and eve discovers those files and compiles them into a runnable app. As the documentation puts it, "the filesystem is the authoring interface" and "the directory is the contract" [3]. On stage, Vercel framed it as "Next.js for agents" [1].

Published as the npm package eve under the Apache-2.0 license, it scaffolds with npx eve@latest init my-agent [3][6]. The framework leans on TypeScript for tools and Markdown for instructions — an instructions.md file alone is a complete, runnable agent, because there's a default model and agent.ts only exists to override it [3][6]. Currently labeled Beta, eve already runs internally at Vercel across what coverage describes as more than a hundred — or by Vercel's own account, hundreds of — agents, six of which were published as open examples [1][6].

Convention over configuration

eve borrows Next.js's "convention over configuration" philosophy and applies it to agents. Each file or subdirectory maps to one capability, discovered at build time rather than registered by hand [5][6].

Tools are TypeScript files in tools/; the filename becomes the tool name, defined with defineTool and a Zod input schema [6].
Skills are Markdown playbooks loaded only when relevant, so guidance stays focused instead of bloating every prompt [6].
Sandbox: every agent runs in its own isolated sandbox by default, in a separate security context, because agent-generated code is treated as untrusted. The backend is a swappable adapter — Vercel Sandbox in production; Docker, microsandbox, or just-bash locally [6].
Approvals are a single needsApproval Boolean on any action for human-in-the-loop control [5][6].
Channels let the same agent run in Slack, Discord, Teams, or the web via one adapter file per surface [6].
Connections handle auth for services like GitHub, Stripe, and Linear without manual token juggling [6].
Subagents live one level down in subagents/ and are delegated to "just like the agent calls a tool" [6].
Schedules run agents automatically with durable continuation [6].

Underneath, each session is a checkpointed durable workflow that resumes after a crash, with OpenTelemetry traces and evals built in and inspectable in the Vercel dashboard [6]. eve works with any model the Vercel AI SDK supports, optionally routed through the AI Gateway for failover [1][6].

Why it matters for practitioners

Vercel's framing is that "agents today are where the web was before frameworks, with everyone hand-rolling the same plumbing and nothing carrying over to the next one" [1]. The value proposition is standardizing that plumbing. MarkTechPost's comparison is blunt: where a DIY stack means a custom loop and manual tool registration, eve is files in a directory; where durability means bespoke retry handling, eve gives a checkpointed durable workflow; where deployment means provisioning infrastructure, eve ships with vercel deploy, unchanged from local [6].

Two consequences stand out. First, agents become Git-native and inspectable — because an agent is just files, it moves through normal Git and CI/CD, and the directory tree shows "what an agent is and does" at a glance [5]. Second, security is a default, not an add-on: sandboxing untrusted code and one-flag approvals tackle two of the most common production pain points up front [6].

The design bet is also distinct. As Techstrong notes, other vendor-agnostic frameworks like LangGraph and Mastra already use the filesystem as a scratchpad; eve's novelty is using the directory structure itself as the workflow router [8][11]. CTO Malte Ubl pitched simplicity as the feature — a "fill in the blanks" model where "the life cycle of the agent is completely orchestrated by the framework" [8].

The lock-in caveat

The biggest tension is open-source framing versus proprietary runtime. Techstrong's analysis is the most skeptical: for production, eve "must run on the proprietary Vercel cloud platform," and running locally "causes the user to miss out on Eve's major features" like durable session state and crash recovery [8]. Documentation for swapping the AI Gateway for alternatives like OpenRouter is, they report, scarce [8]. That sits awkwardly beside Vercel's own positioning as "the anti-vendor-lock-in cloud" [9].

The reconciliation: the framework is Apache-2.0 and runs locally, but the production magic — managed durability, sandbox, gateway — is monetized on Vercel, reportedly starting at $20/month with enterprise rates [8]. eve shipped alongside a broader Agent Stack and governance tooling, including Vercel Connect (short-lived, task-scoped tokens) and Vercel Passport, explicitly framed as a fix for "shadow AI" [4][7].

Bottom line

eve is a concrete, shipped framework betting that agent development will standardize the way web development did — by making the filesystem the authoring interface and turning durability, sandboxing, approvals, channels, and observability into conventions. For practitioners it lowers the cost of getting an agent to production and makes agents reviewable. The caveat is that the most valuable runtime features are tied to Vercel's paid cloud, which critics argue undercuts the "open" label.