This New Protocol Will Change AI Coding Forever (ACP)

Key Concepts

• Agent Client Protocol (ACP): A standard for connecting coding assistants to code editors.
• Zed: A free and open-source code editor with native ACP support.
• Gemini CLI: A coding assistant with direct ACP integration.
• Cloud Code: A coding assistant integrated with Zed through an ACP adapter.
• MCP (Message Passing Communication): A protocol that ACP takes inspiration from.
• Language Server Protocol (LSP): A standard for connecting programming languages to code editors, similar to ACP.
• Archon: An open-source project aiming to be a command center for AI coding assistance, potentially leveraging ACP.
• Fly.io: A developer-first cloud platform for deploying applications, including AI agents.
• JSON RPC: The communication protocol used by ACP, similar to MCP.
• Tool Calls: Actions performed by coding agents, such as reading or modifying files, communicated through ACP.

ACP: The Future of AI Coding

Introduction to ACP

• ACP is presented as a significant advancement, enabling seamless integration between coding assistants and code editors.
• It's a standard, like MCP, designed for connecting any coding assistant to any code editor.
• Zed's implementation is currently in beta but offers a glimpse into the future of AI-assisted coding.

ACP in Action with Zed

• The video demonstrates running both Cloud Code and Gemini CLI threads within Zed.
• Users can switch between coding assistants with a single click.
• ACP provides live updates in the code editor as the coding assistant works.
• Users can accept or reject changes suggested by the AI, improving the development experience.
• ACP allows for the creation of custom agents that can be attached to any editor.

Getting Started with ACP

• To use ACP, install Zed, a free and open-source code editor.
• Gemini CLI has a direct integration with ACP, while Cloud Code uses an adapter.
• Within Zed, the agent panel (toggleable via a button in the bottom right) allows starting new Gemini CLI, Cloud Code, or custom agent threads.
• Authentication for coding assistants is handled within the code editor.
• Users can send requests to the coding assistant, referencing files using @ mentions.
• Native features of coding assistants, such as sub-agents and custom slash commands, are supported.

Beta Integration Caveats

• The integration is acknowledged to be in beta, with occasional issues like hanging during Gemini CLI thread loading.
• Thread history may not be fully functional.

Fundamental Understanding of ACP

• Even if ACP isn't the ultimate standard, future standards will likely resemble it.
• Zed's implementation draws inspiration from MCP and the Language Server Protocol (LSP).

ACP Architecture and Communication

• ACP uses a standard IO connection between the code editor and coding assistants.
• Communication follows a JSON RPC protocol, similar to MCP.
• The process involves creating a session (handshake) and exchanging JSON events for user messages, tool calls, and other interactions.
• JSON events include information like task lists, plain text, and file updates.
• Agent tool calls contain metadata about the file being affected and line numbers.

Building Custom Agents

• ACP enables building custom agents that follow the defined architecture.
• A TypeScript example is provided in Zed's GitHub repository.
• The example demonstrates how to initialize the agent, handle new sessions, and simulate turns.
• Simulated turns involve sending hardcoded JSON messages to mimic agent behavior.
• Custom agents can be added to Zed through the settings, specifying the path to the agent's file.

Cloud Code Adapter

• Cloud Code doesn't have a direct ACP integration, requiring an adapter.
• The adapter converts Cloud Code SDK messages into the JSON format required by ACP.
• The adapter is relatively simple, around 600 lines of code, primarily involving data transformation.

Archon Integration (Proof of Concept)

• ACP facilitates integrating Archon with coding agents like Cloud Code and Gemini CLI.
• A proof of concept demonstrates Archon's ability to invoke coding agents through ACP.
• The vision is to connect a Git repository to Archon, attach a knowledge base, and use MCP servers to provide context to the coding agent.
• The agent can then spin up its own work tree, perform tasks, and return a pull request.
• Archon could manage multiple agents in parallel, providing a user interface to interact with them.
• Users could select specific agents for tasks and monitor their progress.

Fly.io Sponsorship

• Fly.io is a developer-first cloud platform for deploying applications, including AI agents.
• It offers its own infrastructure, resulting in potentially lower costs and faster deployments.
• Deployment is simplified through the Fly.io CLI, which automatically recognizes the application type and configures deployment.

Conclusion

• ACP is presented as a transformative technology with the potential to revolutionize AI coding.
• It enables seamless integration, easy switching between agents, and the creation of custom solutions.
• The video encourages viewers to explore ACP further and anticipate future developments in this area.
• The integration with Archon is a proof of concept and not coming anytime soon to Archon necessarily.