Vercel eve 1.0 RC6: Build Durable AI Agents as Directories of Files

Vercel shipped eve 1.0 RC6 on November 5, and if you’ve been watching the agent framework space, it’s worth pausing on. RC6 isn’t a final release — but it ships a breaking change that tells you exactly where this framework is heading. The concept: every AI agent is a directory of files. No DAG compilation, no YAML configs, no boilerplate registration. Add a capability, add a file. That’s it.

An Agent Is a Directory

Eve borrows the same instinct that made Next.js dominant: file location is configuration. Drop a file in the right folder, the framework picks it up. For agents, that maps out like this:

my-agent/
├── agent.ts         # model config via Vercel AI Gateway
├── instructions.md  # system prompt, plain Markdown, lives in Git
├── tools/
│   └── search.ts    # what the agent can call
├── skills/
│   └── research.md  # context loaded only when needed
├── channels/
│   └── slack.ts     # run the agent inside Slack
└── schedules/
    └── daily.ts     # cron-triggered runs

The tools/skills distinction is cleverer than it looks. Tools are active capabilities — things the agent calls. Skills are passive context, loaded on demand rather than appended to every prompt. That matters for token efficiency: you’re not injecting thousands of tokens of context into every conversation turn, only when the agent actually needs it.

Six Things You Don’t Have to Build

Most agent frameworks are libraries. Eve is infrastructure. It ships six production-critical capabilities out of the box:

• Durable execution. Every session runs as a checkpointed workflow. A crash, a function timeout, or a redeployment mid-turn? The run picks up from the last completed step, not from scratch.
• Sandboxed compute. Agent-generated code runs in Vercel Sandbox, isolated from your application runtime entirely. Runtime-level isolation, not just containers.
• Human-in-the-loop approvals. Any tool can require approval before it fires. The session pauses indefinitely without burning compute. Auto-approval via Jev classifier (clear/caution criteria) keeps routine calls fast.
• Subagents. Delegate to child agents with their own instructions, tools, and execution environments when tasks get too large.
• OpenTelemetry tracing. Built in. No configuration.
• Evals. Scored test suites you run locally or wire into CI.

Vercel isn’t just shipping theory here. They run more than 100 agents internally, and eve-powered agents trigger roughly 29% of all Vercel production deployments. Nearly a third of their production deploys are driven by agents, not humans pushing code.

RC6 Breaking Change: Migrate Before You Upgrade

If you’re on RC5 or earlier and using the agent tool for subagent delegation, RC6 requires a migration. The agentWithModel export is gone.

Before (RC5 and earlier):

export { agentWithModel as default } from "eve/tools/agent"

After (RC6):

import { agent } from "eve/tools/agent"

export default agent({
  models: {
    "openai/gpt-5.5": "Fast tasks and conversation",
    "anthropic/claude-sonnet-4-6": "Complex reasoning and writing"
  }
})

The old API let the model pass any AI Gateway model ID without knowing which ones existed. The new agent({ models }) gives the model a bounded, described menu — it can only pick from what you’ve explicitly listed. This is a deliberate safety improvement: the agent can’t accidentally route to an expensive or unexpected model.

The Lock-In Reality Check

Eve is Apache 2.0 open source, the framework is free, and you can run it yourself. But the honest version: eve’s durable runtime, Vercel Sandbox, and AI Gateway model routing are Vercel-proprietary. Self-hosting exists — you need Postgres for workflow state and Docker for the sandbox — but it’s the second path, not the happy path.

If your team is already deployed on Vercel, that’s not a problem; it’s a feature. If you have multi-cloud requirements or need on-prem, eve is a poor fit by design rather than by omission. That’s a different trade-off than LangGraph (no lock-in, more setup) or CrewAI (Python-first, no built-in durability). Know which you are before you build.

Getting Started

If you want to test the waters:

npx eve@latest init my-agent
cd my-agent
eve dev

Pass --model openai/gpt-5.5 to pick a model at init, or --reasoning high to set reasoning effort. When you’re ready to ship: eve deploy wraps vercel deploy --prod, pulls environment variables, and deploys your agent directory to Vercel Functions and Workflows. The LogRocket guide covers a more complete walkthrough if you want step-by-step detail.

Not 1.0 Yet — But Watch It Closely

The 1.0 GA date hasn’t been announced. RC6 is the sixth release candidate, and the RC6 breaking change signals the API is still being refined — specifically around how agents select and constrain model routing. Treat this as a stable-enough pre-release, not a locked-down stable API.

That said: Vercel uses it in production today at scale, the filesystem abstraction is genuinely novel among TypeScript agent frameworks, and the bundled infrastructure eliminates weeks of custom setup. For teams already on Vercel, it’s the strongest TypeScript agent option in the space right now. Keep the migration guide bookmarked — RC7 is probably coming.