Vercel Open-Sources Eve: Agents Are Just Folders

Vercel has open-sourced Eve, a framework for building and running AI agents that treats each agent as a plain directory of files on disk. It shipped June 17 under the Apache-2.0 license as an npm package, currently in public preview. The pitch is structural rather than flashy: instead of registering tools and capabilities through boilerplate code, you arrange files in folders and the layout itself becomes the agent's contract.

Concretely, an Eve agent is a directory holding an agent.ts (the model and config), an instructions.md (the system prompt), and folders that map to capabilities — tools/, skills/, connections/, channels/, schedules/, subagents/, sandbox/ and lib/. You scaffold one with npx eve@latest init my-agent and deploy it with vercel deploy, unchanged from how it ran locally. Drop a file in the right folder and the agent gains a capability; there's no separate wiring step.

The hard parts come built in

What makes this more than a folder convention is everything Eve bundles. Conversations are durable workflows with checkpointed steps, so a session can pause, survive a crash or a redeploy, and resume where it left off — built on Vercel's open Workflow SDK. Agent-generated code runs in isolated sandboxes (Vercel Sandbox in production, Docker locally). Actions can require human approval, and an agent waiting for sign-off pauses indefinitely without burning compute. File-based 'connections' broker auth to MCP servers and OpenAPI-compatible services like Slack, GitHub, Notion and Salesforce, while 'channels' let the same agent answer over HTTP, Slack, Discord, Telegram and more. Tracing rides on standard OpenTelemetry spans you can export to Datadog, Honeycomb or Jaeger, and evals run locally or in CI.

Why this matters for you

Anyone who has tried to ship a real agent knows the demo is the easy 10%. The other 90% is the plumbing: retries when a step fails, sandboxing untrusted code, pausing for a human to approve a risky action, logging what the agent actually did, and not getting locked into one vendor's runtime. Eve's bet is that all of that should be infrastructure, not something every team reinvents. Vercel says it already runs more than 100 agents in production this way — a data analyst fielding over 30,000 questions a month, a sales agent it pegs at 32x return, a support agent resolving 92% of tickets. Treat those numbers as a vendor's own marketing, but the framework being open and readable on disk is the part that's checkable.