Vercel Ships Eve, Filesystem-First AI Agent Framework with Durable Execution and Sandboxing
Vercel has released Eve, an open-source framework for building, deploying, and operating AI agents in production. The framework organizes agent capabilities as files in a filesystem structure, letting developers define instructions in Markdown, tools as TypeScript files, and reusable knowledge as skills without explicit registration. Eve automatically discovers and exposes these components during build time.
Eve combines production-ready features including durable execution (conversations pause and resume across deployments), sandboxed code execution (via Docker locally or Vercel Sandbox in production), human approval workflows, subagent delegation, built-in tracing, and evaluation suites. The same agent can be deployed across Slack, Discord, Teams, Telegram, GitHub, and HTTP APIs without changing core implementation, with integrations including Salesforce, Notion, Linear, and Snowflake.
Vercel reports Eve powers over 100 internal production agents for analytics, customer support, sales ops, and content review. Architects evaluating agent frameworks should assess Eve against LangGraph, CrewAI, and AutoGen on two dimensions: how its durable-execution model affects cost and latency in long-running workflows, and whether Vercel's integration ecosystem matches your ops stack (Slack/Jira/GitHub vs. other platforms). The filesystem-first design trades discoverability for portability—a significant consideration for multicloud deployments.
Sources
- Primary source
- infoq.com
“Vercel has released Eve, an open-source framework for building, deploying, and operating AI agents in production. The framework provides a filesystem-based project structure that organizes an agent into directories for instructions, tools, skills, subagents, communication channels, and scheduled tasks”