Vercel Chat SDK now runs Claude Managed Agents

Vercel released native support for running Anthropic's Claude Managed Agents through its Chat SDK. The integration moves the entire agent loop, including model calls, tool execution, session state, and sandboxed web research, to Anthropic's servers. Developers get a type-safe handler that works out of the box with Slack, WhatsApp, Discord, Teams, and over 30 other platforms.

The announcement matters because it removes the two biggest headaches of shipping AI agents: managing conversation state and wiring up platform-specific adapters. Both are now handled upstream.

What the integration actually does

Claude Managed Agents runs the full agent loop on Anthropic's infrastructure. That means your server never sees intermediate tool calls or manages the back-and-forth between user prompts and model responses. You send a request, Anthropic orchestrates the work, and you receive streamed tokens as the model writes them.

Chat SDK wraps this in a single handler. Vercel describes the setup as requiring only your Anthropic credentials, with no database, no webhook verification, and no tunnel configuration. The session store lives on Anthropic's side, so features like transcript replay and conversation history read directly from the managed session.

A new Anthropic quickstart template ships alongside the release. It builds a research analyst agent that runs in the browser using Chat SDK's web adapter. The template is meant as a proof of concept. Swapping the adapter moves the same agent to Slack, Teams, or any other supported platform.

Why server-side agent orchestration matters

Running the agent loop client-side or on your own backend creates friction. You manage retries, handle tool call failures, persist conversation state across sessions, and secure any sandboxed execution environments. Each layer is a potential source of bugs.

Anthropic's managed approach centralizes that complexity. The tradeoff is vendor lock-in: your agent logic runs against Claude specifically, and your session data lives on Anthropic's infrastructure. For teams already committed to Claude, that tradeoff is easy. For those hedging across model providers, the abstraction is less portable than it looks.

What the announcement skips

Vercel's changelog is light on operational details. There is no mention of rate limits, session duration caps, or how sandboxed web research handles authentication for private resources. Pricing is also absent. Claude Managed Agents sits behind Anthropic's API billing, but how tool execution and session storage count against quotas is not specified in this release.

Teams evaluating this for production should expect to dig into Anthropic's documentation before committing. The quickstart template is a demo, not a sizing guide.

How to get started

Vercel points developers to the Anthropic quickstart for a working example. The Chat SDK documentation covers adapter configuration for each platform. You need an Anthropic API key with access to Claude Managed Agents, which is currently available through Anthropic's standard developer program.

For teams already using Vercel's edge infrastructure, the integration slots in cleanly. For those on other hosting, the SDK itself is framework-agnostic, though some deployment patterns will require adjustment.