Zed Delta: AI Agent Coding via DeltaDB Version Control

The Core Architecture: DeltaDB

Delta shifts the AI coding paradigm from an editor-centric model to a thread-based workflow. The application is built on DeltaDB, a version control layer that records every operation between commits. Unlike standard Git, which only captures snapshots, DeltaDB assigns a stable identity to every intermediate step and links each code change to the specific agent conversation that generated it. This allows users to trace any line of code back to the original prompt and reasoning process.

Agentic Workflow and Traceability

Delta agents operate within isolated, virtualized work trees, allowing them to experiment, run tests, and commit changes without affecting the user's primary working directory. The agent follows a structured process: it generates design specifications, creates implementation plans, and writes test contracts before executing code changes. Each of these steps is committed to the repository history, ensuring that the thread and the project history remain synchronized. The tool supports collaborative multiplayer sessions via CRDTs, enabling multiple users and agents to co-edit prompts and continue work within the same thread.

Integration and Compatibility

Delta functions as a standalone Rust application compiled to WebAssembly, with a browser-based version available at delta.dev. It maintains compatibility with existing Git repositories and supports a wide range of LLM providers, including OpenAI, Anthropic, and OpenRouter, alongside GitHub Copilot integration. Users can sign in with existing ChatGPT Plus or Pro accounts to access models directly. The platform also implements the Agent Client Protocol (ACP), allowing terminal sessions to sync live into Delta threads, and provides cloud runners to maintain agent progress even when the local application is closed.