Open Agents
lastmod 2026-04-17created 2026-04-171 edits
open-source reference app for building and running coding agents
An open-source reference app released by Vercel Labs that provides an environment for building and running your own background coding agents. It is designed as a three-layer structure: Web → Agent Workflow → Sandbox VM. It is a reference app built on the assumption that you will fork and modify it, and it is not designed to be used as a black box.
Key features
- Agent ≠ sandbox: the agent does not run inside the VM. It interacts with the sandbox from outside through tools (file reading, editing, shell commands, etc.)
- Sandbox Workflow: sandboxes can be created, hibernated, and resumed based on snapshots, and have a lifecycle independent of agent execution
- Durable execution based on Workflow SDK: a chat request starts a workflow run, and an agent turn persists across multiple persisted steps. A run in progress can be picked up again by reconnecting the stream
- GitHub integration: supports repo cloning, branch work, and even automatic commit, push, and PR creation
- One-click Vercel deployment: can be deployed directly to Vercel as a Next.js app
Pros
- Because the agent and sandbox are separated, you can swap models, change the sandbox implementation, and modify agent logic each independently
- Snapshot-based hibernate/resume reclaims resources when idle while preserving state
- The frontend (chat UI) and agent runtime live in the same codebase, keeping integration costs low
Cons
- Tightly coupled to Vercel platform features such as Vercel Sandbox and Workflow SDK, making it hard to port to other infrastructure
- Initial setup requires 10+ environment variables (PostgreSQL, Vercel OAuth, GitHub App, encryption keys, etc.)
- As a Labs project, there is no official guarantee of API stability or long-term maintenance
What sets it apart
Most AI coding agents bundle the agent and the execution environment together. Open Agents adopts, as an explicit design principle, a structure that keeps the agent outside the sandbox and connects them only through tool calls. The VM remains a pure execution environment, and the agent logic runs durably on top of Workflow SDK. It also differs in being offered as a forkable reference implementation rather than a black-box SaaS.
Implications
- The agent-sandbox separation pattern provides a structure in which the model, execution environment, and orchestration can be swapped independently
- Vercel is bundling Sandbox, Workflow SDK, and AI SDK into a single stack to form an agent infrastructure layer
- Teams designing their own coding agents can use this three-layer separation as an architectural reference