Vercel released Open Agents, an open-source reference platform for building and running background coding agents on Vercel’s serverless infrastructure. The project ships a complete stack: web UI, durable agent runtime, sandboxed code execution, and GitHub integration for going from prompt to pull request without a local machine.
What Open Agents Actually Is
Open Agents is a three-layer system. A Next.js web app handles authentication, chat sessions, and streaming UI. An agent workflow layer, backed by Vercel’s Workflow SDK, manages durable multi-step execution that persists across requests. A sandboxed VM handles the actual code: filesystem access, shell commands, git operations, and dev server ports.
The critical design choice: the agent runs outside the sandbox. It reasons and orchestrates from the workflow layer, interacting with the VM through defined tools (file reads, edits, search, shell commands). That separation means sandbox lifecycle, model selection, and agent logic can evolve independently. Sandboxes hibernate after inactivity and resume from snapshots.
The platform supports repo cloning, branch creation, auto-commit, and automatic PR creation after successful runs. Session sharing through read-only links and optional voice input via ElevenLabs transcription round out the feature set.
Why Companies Are Building Custom Agents
Vercel CEO Guillermo Rauch framed the release as a response to a pattern already emerging at major engineering organizations. “Off-the-shelf coding agents don’t perform well with huge monorepos, don’t have your institutional knowledge, integrations, and custom workflows,” Rauch wrote on LinkedIn.
He is not wrong about the trend. Stripe, Ramp, Spotify, and Block have all built internal coding agent systems tailored to their own codebases and workflows, according to Tessl’s analysis of the release. Open Agents gives smaller teams a starting point for the same approach: fork the repo, adapt the tools and context to your codebase, deploy on Vercel.
“The moat of software companies will shift from ‘the code they wrote’ to the ‘means of production’ of that code,” Rauch added. “The alpha is in your factory.”
Where It Fits in the Infrastructure Race
Open Agents occupies a specific lane in what is becoming a crowded agent infrastructure market. Anthropic’s Claude Managed Agents takes the opposite approach: a fully hosted service where Anthropic handles execution, orchestration, and state. AWS launched Agent Registry in public preview earlier this week as a centralized governance catalog. Cloudflare’s Project Think previewed durable agent execution with sub-agents and voice pipelines.
The distinction matters. Managed services trade control for convenience. Open Agents trades convenience for control, requiring teams to deploy, maintain, and customize their own infrastructure. For companies already running on Vercel with large codebases and specific workflow requirements, the tradeoff is straightforward. For everyone else, the operational overhead of running a custom agent platform is real.
The project is designed to be forked, not consumed as a service. That is a deliberate positioning choice by Vercel: infrastructure provider, not agent vendor.