Vercel released Eve, an open-source framework for building, deploying, and operating AI agents in production. The framework uses a filesystem-first architecture where developers organize agent behavior, tools, and skills as files, with automatic discovery at build time. Eve includes production-oriented features like durable execution (conversations persist across failures and deployments), sandboxed code execution, human approval workflows, scheduled tasks via cron, and built-in observability based on OpenTelemetry.
Eve distinguishes itself from competitors like LangGraph, CrewAI, and AutoGen by consolidating infrastructure that teams typically assemble from multiple libraries. Key capabilities include multi-agent orchestration via subagents, integration with external services through Model Context Protocol servers or OpenAPI specs, and deployment across multiple communication channels (Slack, Discord, Teams, Telegram, GitHub, HTTP APIs) from a single codebase. Vercel said Eve is already running over 100 production agents internally.
For architects: Eve's filesystem-driven approach and inclusion of durable execution + evals + sandboxing in a single framework is novel. If you're evaluating agent frameworks, Eve's built-in approval workflows and recovery semantics reduce the 'glue code' burden. Observe community adoption rates and whether the framework's Vercel-centric integrations (Vercel Sandbox, Vercel observability) create lock-in risk for multi-cloud deployments.