Eve – The Framework for Building Agents
6 hours ago
- #Agent Framework
- #AI Development
- #Vercel AI
- An agent in the framework is defined as a directory, where instructions, skills, tools, and configurations are specified in Markdown and TypeScript files.
- The core agent configuration is done via an instructions.md file, describing the agent's role, and optional agent.ts for model selection or runtime configuration.
- Skills are Markdown playbooks that provide focused guidance to the agent only when relevant, reducing prompt clutter.
- Tools are defined in TypeScript files within a tools/ directory, where the filename automatically becomes the tool name without requiring registration.
- Each agent includes an isolated sandbox; a sandbox/sandbox.ts file can customize the backend setup.
- Channels allow the same agent to be deployed across platforms like Slack, Discord, Teams, or the web via channel files.
- Connections handle authentication for services such as GitHub, Stripe, and Linear, enabling tools to call them without managing tokens directly.
- Subagents can be added for specialized tasks, with the main agent delegating work and combining results.
- Schedules enable automatic agent execution for tasks like daily reports using cron jobs, with durable workflows ensuring continuity without active sessions.
- The framework leverages Vercel AI primitives like AI Gateway, Sandboxes, Workflows, and Connect, providing built-in infrastructure for production agents.
- Key features include durable execution with checkpointing, sandboxed compute in isolated VMs, multi-channel delivery, human-in-the-loop approvals, and automated evaluations with test suites.