Vercel's Shocking Eve Launch Finally Makes AI Agents Production-Ready

And honestly? It's about time. We've all built that stateless chatbot that forgets everything on restart. It works for five minutes. It dies in production. Eve doesn't.

An agent is a directory. That's it.

This is the bit that clicks. An Eve agent is just a folder【6911657130774063169†L23-L25】:

agent/
  agent.ts # the model
  instructions.md # who it is
  tools/run_sql.ts # what it can do
  skills/revenue-definitions.md # what it knows
  subagents/investigator/ # who it delegates to
  channels/slack.ts # where it lives
  schedules/monday-summary.ts # when it acts

Each file describes one component. At a glance, the tree tells you what it is, what it does, where it lives, and when it acts【6911657130774063169†L59-L62】. No boilerplate. Eve wires it up. You just focus on what your agent does, not how it runs.

Batteries included, for real

Eve ships with production already built in: durable execution, sandboxed compute, human-in-the-loop approvals, subagents, evals【6911657130774063169†L6-L11】. Not as plugins. Not as "bring your own." It's the default.

1. Durable sessions that survive crashes

Every conversation is a durable workflow with each step checkpointed【6911657130774063169†L126-L129】. Agent waits on people, calls slow systems, runs for hours, days, weeks — doesn't matter. Kill the container mid-run, deploy a new version, it resumes exactly where it stopped. It's built on the open-source Workflow SDK, and it just works.

2. A real sandbox, every time

Code your agents write is untrusted. Eve keeps it out of your app runtime entirely【6911657130774063169†L134-L136】. Every agent gets its own isolated environment for shell commands, scripts, file reads. In production that's Vercel Sandbox. Locally it's Docker, microsandbox, or just-bash. Your agent can grep, run Python, reshape a dataset — safely.

3. Human-in-the-loop that actually pauses

Some actions need a person. Any tool in Eve can require approval, and the agent will pause there and wait, indefinitely if it has to, without consuming any compute【6911657130774063169†L140-L143】. Once approved, it continues right from where it left off. One field: needsApproval: (input) => estimateScanGb(sql) > 50. That's it.

Connect once, run everywhere

Connections are files too. Point at an MCP server or any OpenAPI doc, and Eve discovers the tools, brokers auth, and never shows the model your credentials【6911657130774063169†L145-L148】. At launch you've got Slack, GitHub, Snowflake, Salesforce, Notion, and Linear, plus anything over OAuth or API keys, with Vercel Connect handling token refresh.

And the same agent serves every surface【6911657130774063169†L179-L181】. HTTP API is on by default, with Slack, Discord, Teams, Telegram, Twilio, GitHub, and Linear included. One channel can hand off to another — an incident webhook can open a Slack thread automatically. No more rebuilding the agent for each chat app.

Tools are files. Skills are markdown.

Want to give your agent a capability? Drop in tools/run_sql.ts. It's one typed TypeScript file, filename becomes the tool name【6911657130774063169†L215-L217】. Want to teach it your business logic? Drop in skills/revenue-definitions.md. It loads only when the topic comes up. Eve picks them up at build time. No registry. No wiring.

Need a subagent? It's the same shape, one level down in subagents/, with its own instructions, tools, and sandbox. Parent calls it like a tool, child gets a clean context window, does the work, hands back the result.

Tracing and evals that don't suck

When an agent gets something wrong, Eve gives you a full trace. Every model call, every tool call, inputs, outputs, sandbox commands — all in order【6911657130774063169†L185-L187】. It's standard OpenTelemetry, so it exports to Braintrust, Honeycomb, Datadog, Jaeger. On Vercel, it's in the Agent Runs tab.

And evals are files too. defineEval, write your checks, run locally or in CI. Talk to the agent with eve dev — you get a TUI that shows every skill load, every tool call, every checkpoint, live. It's the dev loop agents have been missing.

Why this matters now

Vercel's been running this internally for years — v0 was built on it. They shipped hundreds of agents and kept rebuilding the same plumbing. Eve is that shape, made into a framework. As they put it: every generation of software earns its abstractions once enough people have built the same thing the hard way【6911657130774063169†L119-L121】.

If you're building invoice bots, onboarding coordinators, sales prospectors, anything with human pauses and cross-system handoffs — this is it. Define the agent as files, let Eve handle durability, sandboxing, approvals, and channels. You write what it does. Eve handles how it runs.

It's Apache 2.0, in public preview now at eve.dev. And yeah, it finally feels like agents grew up.