Sign InOpen Brain
VercelEngineering PostOfficial Source

Build with OpenAI Agents API on Vercel

Vercel pairs OpenAI-managed agent sessions with isolated, persistent Sandboxes, using signed webhooks and queues for reconnection and scaling workers to zero.

Vercel · Sep 10, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The integration runs long-lived, tool-using agents with an **OpenAI-managed loop and session state**, while Vercel hosts the app and attaches each session to an **isolated Sandbox** with persistent files. Signed webhooks and Vercel Queues handle creation and reconnection.

Practical Implication

Builders can separate agent orchestration from execution infrastructure: let OpenAI retain the session while Vercel provides code execution, file access, and a workspace that survives follow-up instructions. The worker architecture can scale to zero.

Agent-Ready Context
The integration runs long-lived, tool-using agents with an **OpenAI-managed loop and session state**, while Vercel hosts the app and attaches each session to an **isolated Sandbox** with persistent files. Signed webhooks and Vercel Queues handle creation and reconnection.

Builders can separate agent orchestration from execution infrastructure: let OpenAI retain the session while Vercel provides code execution, file access, and a workspace that survives follow-up instructions. The worker architecture can scale to zero.

The material does not state persistence duration, sandbox limits, retry semantics, pricing, or recovery behavior when either provider is unavailable, so those boundaries need validation before production use.
Connected Context · Feed7 Judgment

This confirms the emerging brain-versus-hands architecture with a concrete cross-provider integration: managed orchestration and session state are separated from persistent isolated execution. It reduces application-owned lifecycle machinery, but shifts reliability into the seam between providers, making webhook integrity, retries, persistence limits, recovery, and partial-outage behavior essential validation work.

Scaling Managed Agents: Decoupling the brain from the handsAnthropic independently reinforces the same architectural split between agent harness and replaceable sandbox, showing that the brain-versus-hands boundary is not unique to this integration.Cursor Cloud Agents can now run in Vercel SandboxCursor provides a parallel implementation of externally orchestrated agents using Vercel for isolated execution, though its per-request lifecycle contrasts with this target’s persistent workspace and session model.From fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAIThe sandbox-fleet account identifies snapshots, recovery, and placement mechanisms that persistent long-running work may require but this integration leaves unspecified.How Tailscale built a customer-facing model router on AI GatewayTailscale adds an implementation prerequisite beyond orchestration: identity, credential exclusion, retention policy, and sandbox teardown must be designed as part of the execution boundary.
Context Map
infracoding#cloud-agents#sandboxing#agent-reliability
Uncertainty
The material does not state persistence duration, sandbox limits, retry semantics, pricing, or recovery behavior when either provider is unavailable, so those boundaries need validation before production use.