Vercel AI SDK Harness API adds OpenCode and Deep Agents in one interface
Vercel extended the AI SDK Harness API to cover OpenCode and Deep Agents, adding more agent runtimes to the unified interface introduced in AI SDK 7. The change matters because apps can swap supported runtimes without rewriting integration code, though ACP is still awkward for some cloud deployments.

TL;DR
- Vercel added OpenCode and LangChain Deep Agents to the AI SDK Harness API, and vercel_dev's announcement says apps can run them through one interface.
- cramforce's follow-up frames the change as part of a broader push to make popular agent harnesses swappable inside AI SDK.
- The update sits on top of AI SDK 7, where Vercel's release post positioned approvals, durability, and telemetry as the foundation for production agents.
- A notable limitation came from cramforce's ACP reply, which said the current ACP adapter is awkward and not a good fit for cloud deployment use cases.
You can open the announcement link for the runtime support update, browse Vercel's coding agent system post for the broader agent design context, and even see OpenCode running in another surface because charlieholtz's Conductor demo shows it already wired into a multi-model interface.
HarnessAgent abstraction
The core product change is small but useful: HarnessAgent becomes the abstraction layer for more than one runtime. According to vercel_dev's post, developers can run any supported runtime without changing app code, and cramforce's follow-up makes clear that OpenCode and Deep Agents are the next two harnesses added to that interface.
That gives the Harness API a cleaner story than one-off adapters. Instead of binding an app directly to a single agent stack, Vercel is treating harnesses themselves as interchangeable backends.
AI SDK 7 context
The timing matters because this did not ship as an isolated adapter announcement. In Vercel's AI SDK 7 post, the company said the release set the foundation for agents and AI platforms in production, naming approvals, durability, and telemetry as the big primitives.
Vercel's own internal agent work points in the same direction. vercel_dev's engineering post says its coding agents rely on a skill, linters, evals, and an updating loop, which is a reminder that the model is only one piece of the harness.
ACP cloud deployment caveat
The most concrete caveat came from cramforce's reply, which said Vercel already has an ACP adapter but called it awkward, poorly suited to cloud deployment, and better aligned with the "editor on my laptop" use case. The same reply added that the ACP team seems aware of the mismatch and is looking at protocol updates.
That caveat helps explain why Vercel is leaning on a unified Harness API instead of treating every external agent protocol as production-ready. Meanwhile, charlieholtz's Conductor thread shows OpenCode already appearing in other orchestration surfaces, with a demo video of provider switching and code execution, which suggests the broader ecosystem is also converging on runtime portability rather than single-stack lock-in.