Vercel Unveils Eve: A Next.js‑Style Open‑Source Framework for AI Agents Facing Naming Clash
Vercel open‑sources Eve, an agent‑as‑directory framework that bundles production‑grade features such as persistent sessions, sandboxed execution, human‑in‑the‑loop approvals, standardized tool adapters, multi‑channel support and OpenTelemetry observability, while already powering over a hundred internal agents and sparking community debate over its naming.
Background
AI agents grew from less than 3% of Vercel deployments a year ago to 29% of production workloads today, prompting Vercel to open‑source its internal framework Eve, described as “Next.js for agents.”
Agent‑as‑directory design
Eve treats an agent as a single folder following a convention. Example structure:
agent/
agent.ts # model configuration
instructions.md # identity and rules
tools/
run_sql.ts
post_chart.ts
skills/
revenue-definitions.md
subagents/
investigator/
channels/
slack.ts
schedules/
monday-summary.tsOnly agent.ts (one line to select a model) and instructions.md (identity and rules) are required; the framework auto‑generates the runnable agent.
Why build the wheel
Vercel’s blog states that over the past years they built hundreds of internal agents (including v0) and each team repeatedly implemented the same low‑level capabilities—state management, authentication, logging, sandboxing—without reuse. This mirrors the pre‑web‑framework era when teams hand‑crafted HTTP servers and routers.
Built‑in production capabilities
Persistent execution sessions : checkpoints after each step enable pause, crash‑recovery and cross‑deployment continuation, powered by the open‑source Workflow SDK.
Isolated sandbox : generated code runs in a separate sandbox, fully isolated from the application runtime, supporting shell commands and file I/O; Vercel Sandbox is used in production, while local development can use Docker, just‑bash adapters, or custom adapters.
Human‑in‑the‑loop approvals : custom approval rules pause high‑risk tool calls (e.g., large SQL queries) until manual confirmation, conserving compute while waiting.
Standardized tool connections : adapters for any service complying with MCP or OpenAPI, with built‑in integrations for Slack, GitHub, Snowflake, Salesforce, Notion, Linear, handling authentication without exposing credentials.
Multi‑channel operation : a single agent can be connected to Slack, Discord, Teams, Telegram, etc., with cross‑channel conversation flow and HTTP‑triggered alerts that open investigation threads automatically.
Observability and evaluation : each run emits OpenTelemetry‑compatible traces covering model calls, tool invocations and sandbox commands; evaluation test cases can be added to CI for regression testing of prompts or model versions.
Extensibility
Adding a tool, knowledge base, or sub‑agent is as simple as placing a TypeScript file in tools, a Markdown file in skills, or a structured sub‑directory in subagents. All components are auto‑loaded without manual registration.
Development and deployment workflow
Local development: eve dev starts a dev server with a terminal UI showing step‑by‑step execution.
Production deployment: vercel deploy pushes the agent without interrupting running tasks.
Channel integration: eve channels add slack generates the Slack adapter configuration automatically.
Scheduled tasks: add a file with a cron expression under the schedules directory.
Internal use cases
Data analyst agent handling over 30,000 queries per month via Slack, with permissions aligned to user rights.
Autonomous SDR agent that follows up new leads 24/7, costing about $5,000 annually and delivering a 32× return with a single part‑time engineer.
Sales cockpit agent built by RevOps in six weeks without engineers, nearly doubling pipeline coverage.
Support engineer agent that resolves 92% of user tickets automatically, escalating only complex cases.
Routing agent that centralizes internal requests and dispatches them to dedicated agents.
Community reaction
Within an hour of release, developers noted that Eve’s API and core design closely resemble Astro team’s Flue 1.0 Beta announced the day before. Others anticipate competition with frameworks such as Mastra and OpenClaw. Comments ranged from “six directories for one agent is too many” to jokes about a 13‑minute build time similar to Next.js. A naming clash was also mentioned: a prior AI coding platform named “Eva” and the game “EVE Online” sparked tongue‑in‑cheek legal speculation.
Getting started
Eve is available as an open preview; a single command initializes a project in under a minute: npx eve@latest init my-agent Documentation: https://eve.dev/docs
Source code: https://github.com/vercel/eve
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
AI Engineering
Focused on cutting‑edge product and technology information and practical experience sharing in the AI field (large models, MLOps/LLMOps, AI application development, AI infrastructure).
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
