Vercel has released eve, an open-source agent framework for building, deploying, and scaling AI agents in production. Every agent is defined by a file tree: agent.ts for model configuration, instructions.md for the system prompt, and individual TypeScript or markdown files for tools and skills. Eve handles routing, auth brokering, sandboxing, and durable execution through the open-source Workflow SDK, so developers write what the agent does, not the infrastructure underneath it.
The production primitives are not optional add-ons. Every conversation is checkpointed and survives crashes and deploys. Agent-generated code runs in an isolated sandbox separate from the application runtime, backed by Vercel Sandbox in production and Docker or microsandbox locally. Any tool can require human approval, and the agent pauses indefinitely at that step without consuming compute. Connections to Slack, GitHub, Snowflake, Salesforce, Notion, Linear, and any OAuth or MCP-compatible service are each a single file, and the model never sees credentials or endpoint URLs.
The framework analogy Vercel is pushing is deliberate and specific: eve does for agents what Next.js did for the web, turning a folder into a capability by owning the agent loop. The piece worth reading in full is the observability section, which covers per-run traces down to sandbox shell commands, standard OpenTelemetry export to Braintrust, Honeycomb, Datadog, or Jaeger, and a CI-wirable eval system. The local dev tooling, including a UI that lets you watch tool calls and model context in real time, is also documented in detail and is not the afterthought the summary makes it sound like.
[READ ORIGINAL →]