While everyone was pushing Anthropic to release new models, the company dropped this instead.

On August 20, the official ClaudeDevs account announced a new cookbook: pairing Claude Managed Agents with AG-UI. AG-UI is the open-source Agent-User Interaction Protocol from CopilotKit, a standard protocol that enables conversational interactions between agents and user interfaces.

The video shows a demo of a financial assistant. When you ask "If I invest $500 per month with a 7% annual return, how much will I have after 20 years?", the agent doesn't just answer the question—it also renders interactive charts directly in the conversation flow: return curves, growth projections, budget comparisons. All these charts are dynamic, not static screenshots.

The core idea of this cookbook is as follows: Managed Agents handle running the agent loop and containerization, CopilotKit handles rendering the chat interface, and AG-UI acts as the translation layer between the two. Each chat thread corresponds to a managed session, so the agent retains context and supports continuous conversations. Streaming text output, tool calls, and thought processes are all passed to the frontend via AG-UI.

The code is available on GitHub: https://github.com/anthropics/claude-quickstarts/tree/main/managed-agents/copilot-kit-ag-ui

Running the demo is straightforward. If you have Claude Code installed, just run this command directly:

```

claude "walk me through setting up this demo"

```

It will read the CLAUDE.md file in the project and walk you through prerequisites, configuration, and your first conversation. If you don't want to use the CLI, you can do it manually: run `npm install`, `npm run setup`, `npm run dev`, then open `localhost:5173` in your browser.

There are several key details worth noting. In terms of security design: the default environment has no external network access (`networking: limited`, `allowed_hosts` is empty). Bash and file operations are still available, but external API access is blocked. `web_fetch` is disabled to prevent prompt injection attacks. `web_search` remains enabled, as agents need access to real-time exchange rate data. Additionally, the chat endpoint in this demo does not include authentication. You must add your own security protection when deploying, otherwise anyone can use your API quota to run conversations.

In terms of architecture, the translation layer uses the upstream package `@ag-ui/claude-managed-agents`. It maps AG-UI threads to managed sessions, and handles token streaming, tool activity, interruptions, and turn time limits. This repository integrates the adapter into a self-hosted CopilotKit runtime and renders visual interactive tools.

Community reactions have been split. One netizen commented: "The standardization of protocols determines whether agent UI will continue to fragment or converge." Others were unimpressed: "This is the 12th post about managed agents this week. Nobody cares unless the models are working properly again." Another asked: "Can you slow down the video? I can't follow it." These differing perspectives aren't contradictory—they just represent two camps: one focused on infrastructure, the other focused on model performance.

Anthropic's push for protocol standardization at this point indicates the company judges that fragmentation in agent interaction has reached a stage that needs to be addressed. Models can be iterated on, but once a protocol establishes an ecosystem, switching costs become very high. As for whether Anthropic's models are competitive enough, that's a separate topic.