What Eve Is and Why Directory-Based Agents Matter
Vercel’s eve framework is an open source AI agent framework that defines every agent as a directory of files describing its model, prompts, tools, configuration, and runtime behavior, so developers can build and deploy complex agents by organizing code and metadata into a clear, inspectable folder structure instead of hidden pipeline code. Vercel calls eve “Next.js for agents,” signaling a familiar opinionated structure for TypeScript and Markdown projects. Each agent directory includes files for the LLM model (often routed through Vercel’s AI Gateway), a Markdown system prompt, TypeScript tools whose filenames become tool names, and skill.md documents that describe capabilities. Eve compiles this folder into a running agent and orchestrates conversations as durable workflows, using Vercel’s open source Workflow SDK to checkpoint each step. This arrangement aims to make AI agent development less opaque and easier to maintain for teams that are moving past experiments into production deployments.
Inside Eve’s Agent Architecture: A Directory as the Unit of Design
Eve’s directory-centric agent architecture sets it apart from many AI agent frameworks that rely on large configuration files or centralized registries. In eve, the directory itself is the primary unit of design: tools are standalone TypeScript files, the system prompt is a Markdown file, and authentication, channels, and scheduling live alongside them. Nothing needs to be registered in a separate manifest; filenames become tool identifiers, reducing boilerplate and cognitive overhead. Eve also supports skill.md files and MCP servers to connect agents to external systems, plus OpenAPI integrations for more structured services. Once defined, the directory is compiled into an executable agent whose lifecycle is fully orchestrated by the framework. According to Vercel CTO Malte Ubl, the goal is a “fill in the blanks” experience where developers focus on placing the right files in the right locations while eve manages sandboxes, context handling, and workflow orchestration behind the scenes.
Governance, Passport, and the Shadow AI Problem
Beyond agent architecture, eve is part of Vercel’s wider answer to shadow AI in enterprises. As employees build AI-powered apps that auto-select Next.js and Vercel hosting, organizations risk losing oversight of data access and identity. Vercel’s new Passport feature brings eve agents and other applications under centralized identity control using OpenID Connect, integrating with providers such as Okta or Microsoft Entra. Passport, together with Vercel Connect, Enterprise Managed Users, and Bring Your Own Cloud, lets teams tie AI agents to their existing identity and directory systems instead of scattered personal accounts. Each agent runs in an isolated VM sandbox by default, and tools can require human approval before execution, adding another governance layer. According to The Register, Passport was “highly requested” because internal AI projects deployed via Vercel had grown into a form of unmanaged shadow IT, which Vercel now recasts as managed, policy-aware AI.
From Local Dev to Production: Workflows, Observability, and Channels
Eve is designed to collapse the gap between experiment and production. Developers can start an agent locally with a single npx eve dev command and interact through a terminal interface. The same project deploys with vercel deploy, inheriting Vercel’s immutable deployment model so every Git push produces fresh infrastructure. Conversations run as durable workflows, checkpointed through the Workflow SDK so long-running sessions can pause, survive crashes, and resume consistently. Each agent run emits an OpenTelemetry trace that appears in Vercel’s observability dashboard and can be exported to Datadog or Honeycomb, giving teams insight into tool calls and model behavior. Eve also supports multiple communication channels out of the box, including Slack, Discord, Microsoft Teams, Telegram, Twilio, GitHub, and Linear. Vercel reports that agents already trigger around 29 percent of deployments on its platform, up from less than 3 percent a year earlier, signalling growing production use.
Implications for Developer Adoption and the Wider Agent Ecosystem
For developers, eve’s directory-first design and TypeScript focus may lower the barrier to shipping production AI agents compared with more abstract or configuration-heavy frameworks. The clear file layout resembles familiar web stacks, while the lifecycle orchestration, sandboxing, and AI Gateway integration handle infrastructure details behind the scenes. Competitors such as Mastra and LangChain’s LangGraph target similar problems, but eve’s tight coupling with Vercel deployment and observability tools gives it an integrated path from local prototype to monitored production agent. Vercel says it already runs more than 100 internal agents on eve, including a data-analysis agent queried tens of thousands of times a month and a routing agent that delegates work to specialized subagents. For teams standardizing on open source agents, this approach could encourage consistent agent architecture across projects, reduce “one-off” agent scripts, and make governance measures like Passport easier to apply at scale.
