Vercel has launched eve, an open-source, TypeScript-native framework for building, running, and scaling production AI agents, announced on June 17, 2026 at the company's Ship conference. The framework's core idea is unusual: every agent is simply a directory of files that compiles to a durable, production-ready service running on Vercel Functions.
The pitch echoes the company's history with Next.js. "Agents today are where the web was before frameworks," Vercel argues — and eve aims to do for agents what Next.js did for the web by establishing a shared structure that turns scattered building blocks into a single deployable unit.
Key Highlights
- eve is open source, TypeScript-native, and treats each agent as a directory of files rather than a hand-assembled stack.
- Six production capabilities ship by default: durable execution, sandboxed compute, human-in-the-loop approvals, subagents, evals, and OpenTelemetry observability.
- Vercel runs six agents internally on eve, including a data analyst handling more than 30,000 questions a month.
- The company reports its support agent resolves 92% of tickets autonomously.
- Agent-triggered deployments on Vercel have climbed from under 3% a year ago to a majority of all deploys.
Details
In eve, an agent is a folder. A typical agent contains agent.ts for model configuration, instructions.md for the system prompt, and dedicated directories for tools/, skills/, subagents/, channels/, and schedules/. That filesystem-first layout means the structure of the agent on disk maps directly to its capabilities in production — there is no separate orchestration layer to wire up by hand.
The design goal, according to Vercel, is that building an agent should mean defining what it does, not assembling all the moving parts it needs to survive in production. To that end, six capabilities come built in:
- Durable execution — sessions checkpoint at each step, so an agent survives crashes and redeploys without losing state.
- Sandboxed compute — agent-generated code runs in isolated environments.
- Human-in-the-loop approvals — execution can pause and wait for a person to review before continuing.
- Subagents — agents delegate work to child agents.
- Evals — scored test suites validate agent behavior before it ships.
- Observability — OpenTelemetry tracing and structured logging are included out of the box.
Getting started is a single command: npx eve@latest init. Vercel Functions is the primary launch target, with support for additional platforms described as on the way. Alongside eve, Vercel unveiled Vercel Agent for autonomous production monitoring and Vercel for Enterprise Apps and Agents, which adds identity and access controls plus AWS deployment options.
Impact
The most striking part of the announcement is that Vercel is running eve on itself. The company published several agents from its own production fleet: d0, a data analyst fielding more than 30,000 questions a month; a Lead Agent acting as an autonomous sales development rep; Athena, a sales cockpit built in roughly six weeks; and a support agent that, by Vercel's account, resolves 92% of tickets on its own and escalates the rest.
The framing matters for developers weighing the crowded field of agent frameworks. Rather than another library bolted onto an existing app, eve positions agents as first-class deployable services with the same operational guarantees teams expect from web apps — checkpointing, sandboxing, tracing, and gated human review. For enterprises, the human-in-the-loop and observability defaults speak directly to the governance concerns that have slowed agent adoption in regulated environments.
Background
Vercel also shared a metric meant to capture how fast agentic development is moving: agent-triggered deployments on its platform have risen from under 3% a year ago to a majority of all deployments, according to the company. Whether or not that exact trajectory holds, it underlines the bet behind eve — that a growing share of software will be written, deployed, and operated by agents, and that those agents need a real framework rather than glue code.
The release lands in an increasingly competitive market. Open-weight and open-source agent tooling has accelerated through 2026, and eve enters alongside offerings from cloud providers and startups alike. Vercel's differentiator is the tight coupling between the framework and its deployment platform — the same coupling that made Next.js the default way many teams ship React.
What's Next
eve launched in public preview, and Vercel says support for deployment targets beyond Vercel Functions is on the way. For MENA teams building Arabic-language assistants, internal automation, or customer-support agents, the filesystem-first model and built-in evals lower the barrier to shipping agents that can be reviewed, tested, and governed — while the AWS deployment path in the enterprise tier offers a route for organizations with data-residency requirements under frameworks like Saudi Arabia's PDPL and Tunisia's INPDP.
Source: Vercel