Claude Managed Agents Connect to Screens via AG-UI Protocol

Anthropic's frontier lab account has published a new cookbook for Claude Managed Agents. The cookbook shows how to pair Claude Managed Agents with AG-UI, an open protocol created by CopilotKit that standardizes how agents and user interfaces interact. The announcement post summed up the point with the line "Claude Managed Agents can power many different UIs" — meaning a single agent can operate across multiple screens. The attached screenshot showed a "Finance assistant · AG-UI" tab running in a local development environment (localhost:5173), with a user asking, "If I invest $500 a month at a 7% annual return for 20 years, how much will I have?"

Claude Managed Agents is Anthropic's managed agent execution environment for developers, and it has been rapidly expanding its feature set in recent weeks. As covered in Claude Managed Agents Revamps Memory, Domain Controls, and Session Viewer, an August 19 update added the ability to save self-hosted sandbox work to memory, domain allow/block parameters for web_search and web_fetch, and a redesigned session viewer with per-agent lanes and a cost inspector. Earlier, on August 9, the platform gained hard spending caps, an advisor model, regional controls, and automatic skill loading from GitHub repositories. Where those updates focused on hardening the agent's internal capabilities, this AG-UI cookbook takes a different angle, focusing on how the agent communicates with external screens. No matter how capable an agent is on the backend, it's of limited practical use if users have no screen to view results in real time and continue the conversation with follow-up questions. Protocols like AG-UI exist precisely so that developers don't have to rebuild agent integration logic from scratch every time they create a different UI — a chat window, a dashboard, a sidebar, and so on.

With this release, developers can attach Claude Managed Agents to their own web applications while relying on the AG-UI specification to implement input fields, follow-up flows ("Ask a follow-up…"), and response rendering. As in the finance assistant example, this means agent-side logic and interface-side logic can be developed separately when building domain-specific Q&A screens. For teams in Korea building agent-based services, standardizing the contract between the frontend and the agent backend using an open protocol like this could be a way to reduce rework costs whenever the UI needs to be swapped out later.