Hey @luiscaceresd . I ran into the same problem and ended up building a solution that works alongside AI SDK rather than replacing it.

The approach: AI SDK handles the LLM calls and streaming UI (it's great at that), and a separate orchestration layer handles the multi-agent coordination -- task decomposition, dependency ordering, shared memory between agents.

I put together a working Next.js example that shows the pattern:

• Frontend uses useChat from AI SDK for streaming

• Backend API route calls runTeam() from open-multi-agent to orchestrate a researcher + writer agent pair

• The coordinator automatically breaks the goal into tasks, runs them in dependency order, and synthesizes results

• Final output is streamed back via AI SDK's streamText

Architecture looks like:

1
useChat -> POST /api/chat -> runTeam() (orchestration) -> streamText() (streaming) -> browser
2

Working code: https://github.com/JackChen-me/open-multi-agent/tree/main/examples/with-vercel-ai-sdk

The key insight is that these are different layers -- AI SDK is the model/streaming layer, orchestration sits above it. You don't need to choose one or the other.

Caveats: the orchestration phase adds latency (30-60s for coordinator planning + agent execution) before streaming starts, so it's better suited for async workflows than instant chat.

Nice solution. Appreciate you sharing it!

Multi-agent support is not currently available as a built-in feature in the AI SDK. But the project is open source, and contributions are welcome

https://github.com/vercel/ai/blob/main/CONTRIBUTING.md

GitHub

ai/CONTRIBUTING.md at main · vercel/ai

The AI Toolkit for TypeScript. From the creators of Next.js, the AI SDK is a free open-source library for building AI-powered applications and agents - vercel/ai

Ran into the same limitation. The AI SDK handles individual agent loops well but the multi-agent coordination side, where one agent hands off to another with shared context, is not something it covers natively yet.

We tried a few approaches. Google ADK like you mentioned, LangGraph for the state machine side, and a few coordination-focused tools. One of them is teamoffsite.ai which handles the routing and shared thread context between agents as a separate layer. Each agent still runs on whatever SDK or model it uses, the coordination logic just lives outside. Still evaluating whether that separation is the right long-term pattern or whether it makes more sense to wait for the AI SDK to build multi-agent support natively.

Have you looked at the Workflows DevKit for any of this? Jacob Paris posted an evaluator pattern example in a related thread that covers part of the delegation problem, though it is still single-agent at its core.

Heads up, that example link 404s now after a repo restructure. Working link: https://github.com/open-multi-agent/open-multi-agent/tree/main/packages/core/examples/integrations/with-vercel-ai-sdk

GitHub

open-multi-agent/packages/core/examples/integrations/with-vercel-ai-sdk at main · open-multi-agent/open-multi-agent

Self-hosted TypeScript agent runtime with durable approvals and verifiable run records. Own it, approve it, audit it. - open-multi-agent/open-multi-agent

The example I shared earlier is now a one-click Vercel deploy via open-multi-agent (a researcher + writer team running on top of the AI SDK's streaming). https://github.com/open-multi-agent/oma-nextjs-starter