Vercel Chat SDK adds Claude Managed Agents with server-side agent sessions

Vercel has added support for running Claude Managed Agents with its Chat SDK, according to a changelog post on the Vercel blog dated August 27, 2026. The pairing lets developers build chat-driven agent applications where the agent loop — model calls, tool execution, session state and sandboxed web research — runs on the managed service rather than in application code.

What was announced

According to Vercel, Claude Managed Agents handles the agent loop server-side, covering the model, the tools, session state and sandboxed web research. The Chat SDK provides the conversational interface layer on top. Vercel's worked example is a Slack research bot: each Slack thread maps to one persistent agent session, and the agent streams research briefs back to the thread along with their sources.

The changelog entry pitches this as a way to ship an agent app without building the surrounding infrastructure yourself.

What developers get

Vercel highlights four capabilities of the integration:

• Token-by-token streaming. Replies render as the model writes them, delivered over a single streamed response rather than assembled once the turn finishes.
• A live activity feed. Tool calls and model requests are exposed while the turn is still running, so an application can surface a trace of the agent's work inside the chat.
• No database to run. The Managed Agents session stores the conversation itself, and the sidebar, transcript and replay views all read from that session, removing the need for developers to maintain their own server-side state.
• Portability by design. Vercel says swapping a few lines in the handler is enough to move an agent onto Teams, Google Chat, Discord, WhatsApp and more than 30 other platforms.

Why it matters

The announcement is a marker of how the agent application stack is consolidating. Writing an agent into a chat product has historically meant solving a set of unglamorous problems: persisting conversation state in a database, plumbing streaming responses through to the client, and building visibility into what the agent is doing mid-turn. Managed Agents moves all three onto the service side, and the Chat SDK supplies the interface pieces, which shrinks the distance between a prototype and a deployable product.

The single-session-per-thread model also matches how chat platforms actually work, since each conversation thread naturally forms its own context window, and streamed briefs that arrive with their sources fit the expectation that research output should be checkable.

There are trade-offs worth noting. A managed loop means delegating orchestration to the provider, which stays convenient until an application needs custom control over how tools chain together. And while portability across 30-plus chat platforms lowers the cost of switching front ends, it does not by itself reduce dependence on the underlying managed runtime.

Even so, for teams whose goal is a working research or assistant bot inside Slack or Teams, the integration removes most of the boilerplate that made such builds tedious, and it illustrates an emerging division of labour: the provider runs the loop, the platform vendor runs the interface, and developers compose the two.