What Eve Is: An AI Agent Framework Built as Directories
Vercel Eve is an open-source AI agent framework that defines each agent as a directory of files, so its model, tools, prompts, and connections live together in one structured, deployable unit. This directory-based approach treats agents much like web apps in Next.js, which Vercel also maintains, giving developers a familiar mental model. Inside a single Eve directory, TypeScript and Markdown files define the model provider, system prompt, tools, authentication, channels, and schedules, while skill.md files and MCP servers handle external capabilities. Eve compiles this directory into a running agent whose life cycle is fully orchestrated by the framework, from local development to production deployment. By bundling infrastructure, routing, and observability into one project layout, Eve turns agent architecture patterns into something closer to “fill in the blanks” than building a distributed system from scratch.
How Directory-Based Agents Simplify Architecture and Cut Overhead
Treating an AI agent as a directory gives a modular blueprint for agent architecture patterns: one folder per agent, one file per concern. Tools become individual TypeScript files whose filenames are the tool names, so there is no separate registration step. A Markdown system prompt file, model configuration, and channel definitions live alongside those tools, making the whole agent easy to inspect and version. Eve compiles this structure into durable workflows using Vercel’s Workflow SDK, checkpointing each step so long-running sessions can pause, survive crashes, and resume. Because Eve targets Vercel’s platform directly, organizations avoid building their own orchestration layers on top of raw AWS services, which often adds compute overhead and operational complexity. All deployments are immutable, so each Git push creates fresh infrastructure, aligning well with agents that evolve quickly while staying reproducible and easier to reason about in production.
Security, Sandboxing, and Open-Source Access for Developers
Eve bakes security and reliability into the default experience. Each agent runs in its own sandbox on isolated virtual machines, and every tool can be configured to require human approval before it executes, making risky actions easier to govern. Every conversation is a durable workflow that produces an OpenTelemetry trace, surfaced in Vercel’s observability dashboard and exportable to services such as Datadog and Honeycomb. Developers can start agents locally with a terminal command, test behavior with built-in tooling, and deploy using the same vercel deploy flow as web apps. According to Vercel, agents already trigger around 29 percent of deployments on its platform, up from less than 3 percent a year earlier. The Eve framework is available as Vercel Eve open source on GitHub under the Apache 2.0 license, lowering the barrier for teams to build production-ready agents without building infrastructure alone.
Passport and the Fight Against Shadow AI in Enterprises
While Eve simplifies building agents, Vercel’s Passport targets a growing governance headache: shadow AI. Employees increasingly spin up AI-assisted apps that default to Next.js and Vercel hosting, often wired to company data but outside official IT oversight. Passport uses OpenID Connect to put all applications and AI agents in a team behind an identity provider such as Okta or Microsoft Entra, so access is tied to corporate authentication and directory sync. This combines with Vercel Connect, which replaces static secrets with short-lived tokens, and Enterprise Managed Users to keep accounts aligned with HR systems. By bringing Eve-built agents under Passport, enterprises can centralize identity, control who can run which agent, and see what is deployed. That turns Eve’s directory-based agents from potential shadow AI risks into governed assets that fit existing compliance and security models instead of bypassing them.
From Experimentation to Enterprise-Grade AI Agent Architecture
Eve and Passport together push AI agents from experiments toward standard enterprise software. Eve’s opinionated directory layout turns agent architecture patterns into repeatable templates: one directory per agent, predictable files for skills, prompts, channels, and tools, plus durable workflows and traces for every run. Vercel says it already operates more than 100 internal agents on Eve, including a data-analysis agent employees query in Slack tens of thousands of times a month and a routing agent that forwards questions to the right specialist. Combined with Bring Your Own Cloud, which allows Vercel’s platform to run on AWS accounts provisioned by the customer, enterprises can keep infrastructure closer to their standards while benefiting from Eve’s orchestration. For organizations hesitating over trust, cost, and governance, the mix of open-source tooling, sandboxed compute, and shadow AI prevention presents a clearer path to scalable AI agent adoption.
