Alex Hancock argues that bespoke interfaces between AI agents and their client applications limit interoperability. He distinguishes MCP, which lets agents access tools and resources, from Agent Client Protocol, which lets client software assign tasks and receive updates from an agent. A common client interface could let users choose applications without tying each one to a single agent implementation.
Alex Hancock explains that ACP connections negotiate capabilities and support sessions containing user messages, text or media responses, tool-call updates and permission requests. The protocol uses JSON-RPC and allows custom methods. He proposes using shared patterns emerging from those extensions to guide future standardization, and argues that the protocol can serve more than the editors from which it originated.
In the narrated local demonstration, Alex Hancock asks Goose to explain the same single-file HTML project from Zed and a terminal client, showing how one agent-side implementation can serve different interfaces. He then describes an HTTP transport with a WebSocket upgrade that preserves the message semantics, and demonstrates a simple client requesting a poem over the network from a Goose process on his own machine.
Alex Hancock separates the system into client, agent harness, tools and model. He argues that compatible remote transports make it possible to place those components on different machines while retaining shared interfaces. Specialized or personal clients and better user experience are potential benefits he predicts from broader adoption, rather than outcomes established by the demonstrations.
Watch on YouTube
