Vercel Open Agents: Building Autonomous AI Workflows

The shift toward agentic AI continues. Vercel's latest release, Open Agents, is an open-source reference implementation for building autonomous coding agents that run as durable background workflows—not tied to a single request, but capable of persisting, pausing, and resuming across sessions.

Architecture: Three Layers

Open Agents uses a three-layer stack:

1. Web Layer — Authentication, session management, streaming interactions. The UI that humans interact with.
2. Agent Workflow Layer — The actual AI agent running as a durable workflow (not request-bound). This enables tasks to span multiple steps and persist over time.
3. Sandboxed Execution — Isolated virtual machine environments with filesystem access, shell commands, and dev servers. Agents interact with sandboxes via tools, not by executing inside them.

This separation is deliberate. The agent lifecycle and sandbox lifecycle can evolve independently, allowing sandboxes to pause, hibernate, and resume while the agent continues.

What This Enables

• Multi-step workflows — Agents can plan, execute, verify, and iterate across multiple operations without human intervention.
• Persistent state — Sandboxes use snapshot-based restoration, so work persists across session pauses.
• GitHub integration — Clone repos, create branches, commit changes, open pull requests—all automated.
• Streaming outputs — Watch the agent work in real-time, with option to cancel mid-task.
• Voice input — Optional ElevenLabs transcription for hands-free instruction.

The Broader Trend

Open Agents arrives alongside IBM Bob, Vercel's own background agents, and a wave of "vibe coding" tools—where developers describe intent in natural language and AI handles implementation. These tools are accelerating development velocity but raising legitimate questions about code quality, security, and maintenance burden.

An ACM TechBrief warns that while vibe coding tools lack safeguards, 25% of startups are already adopting them faster than enterprises. Vercel's reference implementation is positioned as a governed, transparent alternative—letting teams understand and control how agents work.

For Stewart and Friends

If you're building with this stack, Open Agents is worth exploring. It's designed as a reference, not a finished product, but it demonstrates the architecture that production agentic workflows will need: durability, isolation, and clear separation of concerns.

The platform requires PostgreSQL, GitHub OAuth, and optional Redis—a realistic setup for teams building on Vercel or similar infrastructure.

Source: InfoQ - Vercel Open Agents