Open Agents - An Open-Source Reference App for Building and Running Coding Agents

Vercel Labs has released an open-source reference app that lets you directly build and run background coding agents.

The architecture is designed as a 3-layer structure of Web → Agent Workflow → Sandbox VM. The core idea is separating the agent from the sandbox — the agent doesn't run inside the VM, but instead interacts with the sandbox from outside through tool calls like file read/edit and shell commands.

The sandbox supports snapshot-based creation, hibernate, and resume, and has a lifecycle independent from agent execution. A chat request starts a workflow run, an agent turn continues across multiple persisted steps, and stream reconnection allows picking up an in-progress execution.

GitHub integration is also included.

• Clone repos
• Work on branches
• Automatic commit, push, and PR creation

Deployment is available as a Next.js app with one-click deployment on Vercel.

The advantages are as follows.

• Agent, sandbox, and model swaps can each be changed independently
• Idle resources can be reclaimed via hibernate/resume while preserving state
• Since the frontend and agent runtime share the same codebase, integration cost is low

On the other hand, the limitations are also clear.

• Strongly coupled to Vercel Sandbox, Workflow SDK, etc.
• Requires more than 10 environment variables, including PostgreSQL, Vercel OAuth, GitHub App, and encryption keys
• As a Labs project, there is no official guarantee of API stability or long-term maintenance

Unlike most AI coding agents that bundle the agent and the execution environment together, this reference explicitly adopts a structure that keeps the agent outside the sandbox and connects it only via tool calls. For teams designing their own coding agents, it's a case worth referencing for agent-sandbox separation and durable workflow execution.