Vercel AI Updates: September 11, 2026

1. GitHub Copilot Joined the AI SDK Harness Layer

Vercel. The @ai-sdk/harness-github-copilot adapter lets an application run Copilot through the same HarnessAgent interface it uses for other coding agents, so swapping agents does not require application code changes. The adapter connects via @ai-sdk/harness-acp over the Agent Client Protocol. Supported harnesses now cover Claude Code, Cline, Codex, Cursor, Deep Agents, fx, Grok Build, OpenCode, and Pi alongside Copilot. Source

2. OpenAI’s Agents API Can Now Run With Vercel Sandbox as Its Execution Environment

Vercel. Vercel shipped an integration where OpenAI owns the agent loop and session state while Vercel hosts the application and attaches each session to a Vercel Sandbox for code execution and file access. Sandbox creation and reconnection run through signed OpenAI webhooks and Vercel Queues, each session gets an isolated environment with a workspace that persists files across follow-up instructions, and the architecture scales to zero with no always-on worker. Source

3. Vercel Sandbox Expanded From Four Regions to All Twenty

Vercel. Sandboxes can now run in all 20 Vercel compute regions, up from four, cutting latency to the databases and services they call and letting teams keep workloads inside approved geographies for data residency. iad1 stays the default, region selection is on every plan, and Pro and Enterprise teams can configure ordered failover regions that Vercel tries when the primary region cannot create a sandbox. Active CPU and provisioned memory rates vary by region, and the SDK or CLI must be updated before selecting a newly supported one. Source

4. Every Sandbox Now Gets 64 GB of Storage

Vercel. Storage per sandbox doubled from 32 GB to 64 GB, covering sandboxes created from a Vercel Managed Image, a custom image, or the deprecated runtime property. The stated targets are large repositories, heavy dependency trees, build artifacts, data workloads that spill to disk, and storage-hungry agent tasks. Source

5. Tako Search Is Free on AI Gateway Through September 30

Vercel. Tako Search is available at no cost on AI Gateway until September 30, after which it bills at standard rates. It lets models search Tako’s curated data plus the live web, filter results by domain or publication date, and return answers with citations and visualizations. Add gateway.tools.takoSearch() to a generateText or streamText call; no separate Tako account or API key is needed, and the same setup works across any model on the Gateway. Source

6. FastAPI Frontends and Static Files Get Promoted to the CDN at Build Time

Vercel. Paths served with app.frontend() or StaticFiles are now promoted to the Vercel CDN during the build and served without invoking the function. FastAPI’s declaration-order route precedence is preserved, and promoted source directories stay in the function bundle so the app can still read them at runtime unless tool.vercel.fastapi.static.exclude = true is set. Frontends behind dependencies or static files behind middleware stay on the function, since the CDN cannot run those checks, overridable with tool.vercel.fastapi.static.cdn. Source

7. Vercel Cut P99 CDN Metadata Lookup Latency by 91% by Batching Path Fetches

Vercel. Vercel’s CDN executes over 80 million routing instructions per second on average, and part of that is resolving which paths exist and how to serve them. The old design fetched and cached metadata one path at a time, which meant deployments with hundreds of thousands of paths paid a recurring cache-miss cost on every new deploy. Grouping paths so one fetch populates the cache for many subsequent lookups cut P99 metadata lookup latency by 91%, with production experiments used to find the point where fetching more stopped paying for itself. Source