The Filesystem-First Architecture

Eve shifts the paradigm of agent development from complex, hand-rolled codebases to a structured, filesystem-first approach. By defining an agent as a directory on disk, the folder structure itself serves as the contract for the agent's capabilities. This design allows developers to add functionality—such as tools, skills, or communication channels—simply by adding files to specific subdirectories. The framework automatically discovers and wires these components at build time, removing the need for manual registration boilerplate.

Standardized Agent Contract

The framework enforces a predictable shape for agents, which Vercel claims is based on the architecture they use for over 100 production agents. The required structure includes:

  • agent.ts: Defines the model and runtime configuration.
  • instructions.md: Acts as the system prompt for every model call.
  • tools/: TypeScript files using Zod schemas for input validation.
  • skills/: Markdown files containing knowledge loaded dynamically.
  • connections/: Secure links to MCP servers and OpenAPI APIs.
  • channels/: Adapters for external interfaces like Slack or HTTP.
  • schedules/: Cron-based triggers for autonomous actions.

Production-Ready Capabilities

Eve aims to solve the "plumbing" problem inherent in moving agents from local prototypes to production. It provides built-in support for:

  • Durability: Checkpointed workflows that handle retries per session.
  • Approvals: A native needsApproval field in tool definitions, allowing for programmatic human-in-the-loop gates.
  • Observability: Native integration with OpenTelemetry for tracing and evaluation.
  • Deployment: Seamless integration with Vercel's infrastructure, allowing developers to use vercel deploy to ship agents without changing code.

Local development is supported via eve dev, which provides an interactive terminal UI, while eve eval handles test suite execution. By standardizing these operational concerns, Eve allows developers to focus on agent behavior rather than infrastructure management.