Summary
Alex Hancock, software engineer at Block (parent of Cash App and Square) and a maintainer of the Model Context Protocol (MCP), takes the stage at AI Engineer to introduce ACP — the Agent Client Protocol — an emerging open standard for how client software communicates with AI agent harnesses. His argument: while MCP has solved the problem of agents calling tools in external systems, there is still no standard for how a client tells an agent what to work on and receives progress updates back.
ACP, which originated from collaboration between Zed and JetBrains, uses JSON-RPC and supports sessions, streaming text and media responses, tool call notifications, and permission request flows. Critically, it is designed to be extensible — teams can add custom methods with an underscore prefix, with the expectation that patterns that emerge across multiple harness projects will eventually be standardized. Hancock demos ACP using Goose (Block’s open-source agent, now under the Linux Foundation) controlled simultaneously from both the Zed text editor and a separate terminal client, showing the same harness serving two different clients without any changes on the agent side.
For platform teams and IDE developers, ACP represents a potential inflection point: a single high-quality client implementation could control any compliant agent harness, much as HTTP enabled one browser to reach any web server. Poolside AI is among the early adopters shown in the demo.
📺 Source: AI Engineer · Published September 09, 2026
🏷️ Format: Keynote Launch
