Vercel Open Agents: Why Big Tech Is Building Their Own AI Coding Factories

Stripe ships 1,300 PRs per week with zero human-written code. Ramp’s AI writes over half of all merged pull requests. And now Vercel just gave everyone the blueprint to build the same thing.

Table of Contents

The Rise of AI Software Factories

Something fundamental is shifting in software engineering.

It’s not about AI writing a few lines of code for you. It’s about entire autonomous systems that take a task description, write the code, run tests, open a pull request, and wait for a human to hit “merge.”

Companies like Stripe, Ramp, Spotify, and Block aren’t just using AI coding tools — they’re building their own AI coding factories from scratch. Internal platforms. Custom agents. Proprietary workflows.

Why? Because off-the-shelf tools can’t handle what they need.

And now, Vercel just open-sourced the blueprint.

What Is Vercel Open Agents?

Open Agents is an open-source reference platform for building cloud-based coding agents. Think of it as a production-ready template that you can fork, customize, and deploy.

It’s not a chatbot. It’s not a VS Code extension. It’s a full-stack infrastructure for autonomous agents that:

• Clone repositories
• Create branches
• Write and edit code
• Run tests
• Commit changes
• Open pull requests

All of this happens in the cloud — no local machine required.

The Tech Stack

Open Agents runs on Vercel’s agentic infrastructure:

Why Companies Build Their Own Agents

This is the question everyone’s asking. GitHub Copilot exists. Cursor exists. Claude Code exists. Why would a company spend engineering resources building their own agent?

Two reasons — one technical, one strategic.

1. Technical: Off-the-Shelf Doesn’t Scale

Generic coding agents hit a wall with enterprise codebases:

Stripe’s codebase is Ruby with Sorbet typing. Spotify has custom migration pipelines. Ramp has proprietary security requirements. No general-purpose agent handles any of this out of the box.

2. Strategic: The Factory Is the Moat

Here’s the insight that’s reshaping the industry:

The competitive advantage of software companies is shifting from “the code they wrote” to “the means of production” of that code.

If AI can write code, then code itself becomes a commodity. What remains valuable is:

• How fast you can ship
• How reliably your agents produce correct code
• How deeply your automation integrates with your specific domain

The company with the best AI coding factory ships faster, iterates faster, and outpaces competitors — regardless of team size.

The alpha is in the factory, not the code.

Who’s Already Doing This?

Stripe — Minions

Stripe’s Minions are fully autonomous coding agents that produce 1,300+ PRs per week with zero human-written code. A single Slack message triggers an end-to-end workflow: understand the task, write the code, run tests, open a PR.

Key details:

• Built on a forked version of Block’s Goose agent
• Uses a blueprint model combining deterministic steps with AI-driven reasoning
• Runs in isolated devbox cloud environments
• Deeply integrated with Stripe’s CI/CD pipeline

Ramp — Inspect

Ramp’s Inspect agent writes over 50% of all merged PRs across frontend and backend repos. It handles code reviews, security patches, dependency updates, and more — triggered by events, not just human commands.

Spotify — Background Agents

Spotify deployed background coding agents that have produced 1,500+ merged PRs, cutting migration time by 60-90%. Their agents are event-driven, accelerating the entire software delivery system.

Block — Goose

Block (formerly Square) created Goose, an open-source agent harness that became the foundation for Stripe’s Minions. Goose handles agent orchestration — combining workflow steps with LLM-powered reasoning in a robust, extensible framework.

Open Agents Architecture: How It Works

Open Agents uses a clean three-layer architecture:

┌─────────────────────────────────────┐
│           Web Layer                 │
│  Auth · Sessions · Real-time UI     │
├─────────────────────────────────────┤
│          Agent Layer                │
│  Durable Workflows · Multi-model    │
│  Task Delegation · Checkpointing    │
├─────────────────────────────────────┤
│         Sandbox Layer               │
│  Isolated microVMs · File System    │
│  Shell · Git · Network · Dev Server │
└─────────────────────────────────────┘

Web Layer

Handles authentication (OAuth with Vercel), session management, and real-time streaming UI for monitoring agent activity.

Agent Layer

The brain. Uses Vercel’s Workflow SDK for durable execution — agents can hibernate on inactivity, checkpoint progress, and resume instantly after failures or restarts. Supports multi-agent architectures with explorer and executor subagents running in parallel.

Sandbox Layer

Ephemeral, isolated microVMs with full file system, shell, Git, and network access. Agents execute code in a secure environment that can be snapshotted and restored.

Key Features That Matter

Autonomous GitHub Integration

Agents clone repos, create branches, commit code, and open PRs — all without human intervention. The entire git workflow is automated.

Durable Workflows

Unlike simple API calls, agent tasks can run for minutes or hours. The Workflow SDK ensures they survive crashes, restarts, and network issues by checkpointing state automatically.

Multi-Model Orchestration

Through the AI Gateway, agents can use multiple foundation models (OpenAI, Anthropic, Google) in a single workflow — picking the best model for each subtask.

Sandboxed Execution

Every agent runs in an isolated microVM. No risk to production systems. Full shell access for testing, building, and running code.

Open Source (MIT License)

Fork it. Customize it. Deploy it. No vendor lock-in. The entire codebase is yours to modify.

How Open Agents Compares to Copilot and Cursor

The key distinction: Copilot and Cursor augment developers. Open Agents replaces entire workflows.

Getting Started with Open Agents

Prerequisites

• Vercel account
• GitHub OAuth app
• PostgreSQL database (for sessions)
• API keys for at least one AI model

Quick Start

# Clone the repository
git clone https://github.com/vercel-labs/open-agents.git
cd open-agents

# Install dependencies
pnpm install

# Configure environment variables
cp .env.example .env.local
# Edit .env.local with your credentials

# Deploy to Vercel
vercel deploy

What You Can Customize

• Agent behavior — modify prompts, tool sets, and decision logic
• Model selection — choose which AI models handle which tasks
• Integration points — connect to your CI/CD, Slack, Jira, or any internal tool
• Security policies — define what agents can and cannot do
• Workflow templates — create reusable task blueprints

The Bigger Picture: Code as a Commodity

We’re entering an era where writing code is no longer the bottleneck. The bottleneck is:

1. Knowing what to build — product vision, user needs, business strategy
2. Orchestrating how it gets built — workflows, quality gates, compliance
3. Operating the factory — the infrastructure that turns intent into shipped software

Open Agents is significant not because it’s another AI tool — but because it democratizes the factory model that was previously only available to companies with massive engineering budgets.

Stripe spent months building Minions. Ramp invested heavily in Inspect. Spotify built custom migration agents.

Now, with Open Agents, a 10-person startup can deploy the same architecture in a weekend.

The playing field just got a lot more level.

Frequently Asked Questions (FAQ)

Is Open Agents free?

The platform itself is open source (MIT license) and free. You’ll pay for infrastructure costs — Vercel hosting, database, and AI model API calls.

Does it only work with Vercel?

The codebase is optimized for Vercel’s infrastructure (Fluid, Workflow SDK, Sandbox), but since it’s open source, it can be adapted for other platforms.

Can I use it with my private repositories?

Yes. Open Agents integrates with GitHub via OAuth and can access any repository you authorize, including private ones.

How is this different from GitHub Copilot Workspace?

Copilot Workspace assists developers within GitHub’s UI. Open Agents is a self-hosted, fully autonomous platform that you control and customize end-to-end.

Is it production-ready?

Vercel calls it a “reference platform” — it’s designed to be forked and customized for your specific needs rather than used as-is in production.

What models does it support?

Through the AI Gateway, it supports OpenAI, Anthropic, Google, and any other provider you configure. You can use different models for different tasks within the same workflow.

Vercel Open Agents is available now at open-agents.dev and on GitHub. The era of AI software factories is no longer limited to Big Tech.