What changed
Vercel's changelog says the AI SDK harness layer can now work with ACP-compatible harnesses, extending the harness architecture it introduced earlier this summer. That earlier HarnessAgent launch gave AI SDK one interface for running external agent runtimes such as Codex, Claude Code, and Pi without rewriting the surrounding app.
The new piece is interoperability. The official AI SDK ACP provider docs describe a path for connecting agents that speak the Agent Client Protocol, while the ACP introduction frames the protocol as a common way for clients to discover and use agent capabilities. In practical terms, Vercel is pushing AI SDK beyond "call a model" and closer to "plug in whichever full agent runtime fits this workflow."
Why it matters
This is a useful product shift for builders because coding agents are starting to look like replaceable infrastructure rather than one-off integrations. If your application already uses AI SDK for chat, tool calls, streaming, or telemetry, ACP compatibility means you can point the same product surface at more than one external agent runtime instead of baking one vendor's harness deeply into the app.
That matters for internal developer tools, repository copilots, and workflow automations where teams may want to compare Codex-style flows, Claude Code-style flows, or newer protocol-compatible agents without reworking the whole interface. The stronger the protocol layer becomes, the more the app can treat agent choice as configuration instead of architecture.
The broader signal
The timing matters too. On the same August 13 changelog page, Vercel also announced that Grok Build is now available in the AI SDK harness layer. That makes the ACP update look less like a minor adapter and more like a platform move: Vercel wants AI SDK to be the control surface that sits above multiple agent runtimes.
For portfolio and product readers, the takeaway is straightforward: agent products are becoming more modular. The winning teams may not be the ones that bet perfectly on a single agent runtime. They may be the ones that design their product layer so they can swap, evaluate, and govern multiple agent backends as the ecosystem keeps moving.