Eve is a specialized framework developed by Vercel for building AI-powered agents. It follows a modular, directory-based structure similar to how frameworks like Next.js handle web applications. By utilizing Markdown for instructions and skills, and TypeScript for defining tools, developers can create robust, durable agents that scale automatically. Eve is built on top of Vercel’s Agent Stack, integrating seamlessly with AI Gateway, Workflows, Vercel Sandbox, and Vercel Connect, allowing developers to focus on logic rather than underlying infrastructure.

The framework compiles an agent directory into an executable system that manages durable workflows, connects to various communication channels, and executes tools in isolated environments. It simplifies the development of complex agents by handling state persistence, event streaming, and subagent delegation, ensuring that agents remain operational and responsive without requiring active sessions or constant manual oversight.

Some of the key features are:

• Durable Execution: Workflows survive crashes and restarts by checkpointing steps, allowing agents to park when waiting and resume seamlessly upon delivery.
• Sandboxed Compute: Agents run in isolated VMs on demand, providing secure file system access and bash execution while maintaining full isolation.
• Multi-channel Connectivity: A single codebase can be deployed across various channels including Slack, Discord, Microsoft Teams, and web chat.
• Subagent Delegation: Developers can build specialized subagents with their own unique prompts, tools, and sandboxes that the main agent can delegate tasks to.
• Reusable Skills: Markdown-based playbooks can be loaded when relevant, providing the agent with focused guidance without burdening every individual prompt.
• Tool Definitions: Tools are defined in TypeScript files where the filename determines the tool name, removing the need for explicit registration.
• Scheduled Operations: Agents can run on defined schedules for tasks like daily reports or digests, ensuring work continues even without user interaction.
• Automated Evaluations: Developers can define test suites with scoring rubrics to run evaluations on every deployment or on a recurring schedule.

To use Eve, a developer initializes a new project using the CLI, which generates the required directory structure. Within the agent folder, they define identity and behavior in an instructions.md file, configure the runtime or models in agent.ts, and implement functions in tools/. Once defined, the framework takes over by wiring up the durable workflows, connecting the necessary channels, and managing the sandbox environment during execution. The framework handles the complexities of authentication via Vercel Connect and model routing through the AI Gateway, ensuring production-ready performance.

Some common use cases include:

• Customer Support Agents: Building agents that handle customer queries, check order statuses, or escalate complex issues to human agents using approval gates.
• Scheduled Reporting: Creating agents that pull data from various sources like GitHub, Stripe, or Linear to generate and send daily or weekly summaries to team members.
• Internal Tool Automation: Developing agents that interact with internal databases or APIs to automate mundane tasks like data integrity checks or SQL query execution.
• AI Research Assistants: Constructing agents that investigate novel or ambiguous topics by gathering evidence across multiple sources and synthesizing the findings into a coherent answer.