Vercel Eve: open-source agent-as-directory framework ships

Vercel Eve is the most operationally complete agent framework shipping today — backed by a fleet of over 100 production agents, one of which returns 32x its annual cost. But it is not platform-neutral. Every production capability it ships — durable execution, sandboxed compute, credential brokering, model routing — runs on Vercel infrastructure. If you are already on Vercel, Eve is the fastest path from prototype to production agent. If you are not, the framework is free but the runtime is not.

What happened

Vercel launched Eve on June 17, 2026 at its Ship 26 conference in London: an Apache 2.0, TypeScript-native agent framework where every AI agent is a directory of files on disk. The current version at launch is eve@0.11.4, available through npx eve@latest init. The framework is in public beta, and Vercel warns that APIs may change before general availability.

An Eve agent is defined by its directory structure, not by configuration objects. At minimum, an agent requires two files: agent/agent.ts for model configuration and agent/instructions.md for the system prompt. Capabilities are added by convention: a TypeScript file in agent/tools/ becomes a callable tool where the filename is the name; a Markdown file in agent/skills/ becomes a playbook loaded on demand; files in agent/channels/ connect the agent to Slack, Discord, Teams, Telegram, Twilio, GitHub, or Linear with no webhook plumbing.

Eve ships with six production capabilities built in: durable execution via Vercel's Workflow SDK (session state checkpointed at every step and deterministically replayed after crashes), sandboxed compute (Vercel Sandbox in production, Docker or bash locally), human-in-the-loop approvals (a single flag on any tool pauses the agent until approval arrives), subagents, OpenTelemetry tracing, and a built-in evals system. Model requests route through Vercel's AI Gateway, which handles provider fallbacks transparently across OpenAI, Anthropic, Google Gemini, and open-weight models.

Vercel disclosed that agents now trigger more than half of all commits on its platform, up from under 3% six months ago. Its own fleet includes d0, a data-analysis agent handling over 30,000 Slack queries a month with automatic data-permission enforcement; the Lead Agent, an autonomous SDR that costs roughly $5,000 a year to run and returns 32 times that; Vertex, a support agent resolving 92% of tickets without escalation; Athena, a sales cockpit built by the RevOps team in six weeks; and V, a routing agent that makes hundreds of specialist agents feel like a single system (TechTimes, 2026).

Why Vercel Eve matters

The agent-framework space is fragmented. Most frameworks give you the AI logic — model orchestration, tool calling, prompt management — and leave all infrastructure concerns to the developer. Durable execution, sandboxed code, credential management, observability, and deployment pipelines are hand-rolled separately for every agent. Vercel's argument is that every one of its internal teams had to build these pieces from scratch, and none of it carried over between projects.

Eve ships infrastructure as a first-class concern. This is what production agents actually need, and the internal fleet numbers — 100+ agents, 32x ROI on one, 92% ticket resolution on another — are real proof that the model works at scale. The directory-as-agent architecture is genuinely novel: it makes agents version-controllable, composable, and inspectable in a way that object-in-memory frameworks do not.

The tradeoff is platform coupling. Durable execution requires Vercel Workflow. Sandboxed compute uses Vercel Sandbox. Credential brokering flows through Vercel Connect. Each of these has a local adapter (Docker, bash, microsandbox), but porting the production runtime to a non-Vercel host means replacing every adapter — technically possible under Apache 2.0, but operationally significant. This is the Next.js playbook: open-source the framework, monetize the runtime, and let distribution — Vercel serves over six million developers — do the rest (TechTimes, 2026).

Teams already running agent workflows on tools like n8n (which we rate Recommended for self-hosted AI automation) should weigh whether Eve's integrated deployment story justifies the platform switch.

For the broader ecosystem, Eve raises the bar on what an agent framework should ship out of the box. Platform-neutral alternatives like Mastra (TypeScript, v1.0, host-agnostic), LangGraph (Python-first), and Inngest AgentKit (TypeScript, durable execution) now have a harder benchmark to meet. The cloud providers — Amazon Bedrock AgentCore, Google Vertex AI Agent Engine, Microsoft Agent Framework — offer managed runtimes that accept any framework, but none yet match Eve's integrated local-to-production story.

What changes for you

If you are on Vercel, Eve eliminates months of infrastructure work. The local development experience runs with Docker alone, and the deploy path to Vercel Functions is a single step. Start prototyping with npx eve@latest init — the framework is free, and the sandboxed compute gives you safe code execution from day one.

If you are not on Vercel, weigh the platform dependency carefully. The Apache 2.0 license lets you fork and run Eve anywhere, but self-hosting the durable execution layer and sandboxed compute at scale is untested. Mastra (platform-agnostic, TypeScript-native, reached v1.0 in January 2026) and LangGraph (Python-first, most established) are the strongest alternatives for teams that need runtime portability.

For a broader view of the agent-framework landscape, see our coverage of Sakana Fugu, a different approach to multi-agent orchestration.

If you are evaluating for production, note that Eve is in public beta. Vercel states that APIs may change before general availability. The ROI numbers from Vercel's internal fleet are real, but they reflect a team deeply integrated with the platform — your mileage will vary.

FAQ

Is Eve truly open source? Yes. The framework is Apache 2.0 — you can fork it, modify it, and run it locally with Docker. The production runtime capabilities (durable execution, sandboxed compute, observability) depend on Vercel infrastructure, but the local development experience is fully functional without a Vercel account.

How does Eve compare to LangChain or LangGraph? LangChain's LangGraph is Python-first and the most established agent framework; it supports durable execution and has LangSmith for observability, but requires teams to manage their own hosting infrastructure. Eve is TypeScript-native and bundles deployment as a first-class concern. For TypeScript teams on Vercel, Eve eliminates the operational work that keeps agent prototypes from reaching production. For teams on other platforms or in Python, LangGraph offers more flexibility at the cost of more setup.

When should I adopt Eve in production? Eve is in public beta (v0.11.4). Vercel explicitly states APIs and behavior may change before GA. If you are comfortable with the same dependency risk Vercel's own teams manage — and can tolerate API churn — the framework is used in production by Vercel itself at scale. Otherwise, wait for a stable release or evaluate platform-neutral alternatives.