Vercel has launched Open Agents, an open-source app that enables the creation and execution of background coding agents. It provides a complete stack for developers to run independent coding workflows without relying on local machines.
Open Agents is structured as a three-layer system: a web interface, an agent workflow layer, and a sandboxed execution environment. The web layer manages authentication, sessions, and streaming interactions. The agent itself runs as a durable workflow, while code execution happens inside isolated virtual machine sandboxes that provide filesystem access, shell commands, and development servers.
A central architectural decision is the separation between the agent and the sandbox. The agent does not execute directly inside the virtual machine but interacts with it through tools such as file operations, search, and shell commands. This design allows the agent lifecycle and sandbox lifecycle to evolve independently. It also enables workflows to persist beyond a single request, with sandboxes able to pause, hibernate, and resume as needed.
The platform supports multi-step execution, streaming outputs, and task cancellation. Developers can clone repositories, create branches, and automate commits and pull requests with GitHub integration. Sessions can be shared via read-only links, and voice input is available through ElevenLabs transcription.
From a runtime perspective, Open Agents relies on durable workflows rather than request-based execution. Each interaction triggers a workflow run that can span multiple steps and persist over time. Active sessions can be resumed by reconnecting to the workflow stream, while sandboxes use snapshot-based state to restore environments after inactivity.
The project needs several infrastructure components to work, including a PostgreSQL database for managing state and authentication through OAuth. It also involves GitHub integration for accessing repositories and automating tasks like creating pull requests. Optional tools like Redis or key-value storage can be added for caching data.
Open Agents is intended as a reference implementation rather than a finished product. The repository is designed to be forked and adapted, giving developers visibility into how agent workflows, sandbox orchestration, and integrations are wired together. Deployment is supported on Vercel, with environment variables and services configured manually to expose the underlying system design.
The separation between agent logic and execution environment has drawn mixed reactions. Michiel Voortman commented:
You mentioned that the separation between VM and agent is the main point of the project.
I think this slows down agentic development in the mid to long run. It probably scales better, but you limit the freedom of the agent and thus his capabilities and ‘creativity’. I think you need to start with giving each agent a computer. Then you give them 10, 100, 1000 computers as the models become powerful and the handoffs become better.
The release signals a shift towards viewing coding agents as long-running systems instead of mere request-bound tools. By merging workflow orchestration with isolated execution environments, Open Agents enables the development of continuous agents that can interact with external systems and autonomously manage code changes.