Delta is a collaborative, agent-based development environment created by the team behind the Zed editor. It fundamentally changes how software is built and reviewed by keeping conversations, context, and code changes tightly integrated within a single, persistent workspace. By utilizing a custom version control layer known as DeltaDB, the platform records every edit made by both developers and agents, allowing teams to review decisions and iterate on code with full historical context. Unlike traditional git-based workflows that only capture snapshots at commit time, Delta provides a granular stream of modifications that remains linked to the original agentic conversation. This allows users to understand the rationale behind specific changes, trace the evolution of a feature, and easily revert to previous states without managing messy commit histories. The environment is designed to be highly interactive, enabling teammates to jump into threads, comment on specific lines of code, and collaborate alongside AI agents in real time. Delta manages project state through Delta worktrees, providing isolated environments for agent activity while ensuring consistency with local project folders. Whether working locally or in a shared cloud-based environment, Delta keeps the conversation and the project in perfect synchronization. The platform supports various workflows, allowing users to adopt existing git repositories while benefiting from the structured, agent-centric record-keeping of DeltaDB. By connecting developers, teammates, and AI agents in a shared interface, Delta aims to provide a more intuitive and meaningful way to manage complex software development processes.

Some of the key features are:

• Thread-Based Context: Conversations, code edits, and project history are kept together in a single, shareable artifact.
• DeltaDB Versioning: Tracks every granular change made by agents and users continuously, independent of git commits.
• Multiplayer Collaboration: Enables real-time teamwork by allowing multiple users to join threads to review, comment, and guide agent activity.
• Isolated Workspaces: Utilizes Delta worktrees to provide dedicated, parallel environments for agent tasks without disturbing local project files.
• Agent-Agnostic Interaction: Integrates seamlessly with external LLM providers and hosted models, allowing flexibility in which intelligence drives the work.
• Direct Git Integration: Compatible with existing git repositories, with support for Jujutsu and automated syncing between Delta-managed checkouts and local environments.
• Contextual Review: Feedback on code is anchored to the specific conversation and edit, ensuring that context remains preserved as the codebase evolves.

Delta operates by organizing development work into threads where a user engages with an AI agent. When a project is added to a thread, Delta creates an isolated Delta worktree—a replica of the project's state that is fully synchronized with DeltaDB. The agent performs operations within this worktree, and the resulting changes, along with the associated conversation, are stored and versioned in the database. Users interact with this environment through the Delta desktop application or via web interfaces, where they can oversee the agent's progress, provide guidance, or comment directly on the work. Once a task is complete, the changes can be synchronized back into the user's primary git checkout, ensuring a clean and manageable integration process.

Some common use cases include:

• Iterative Feature Development: Using an agent to implement a specific feature while keeping the step-by-step reasoning and discussions for later review.
• Code Review and Refinement: Inviting a teammate into a thread to review agent-generated code and providing contextual feedback directly on the changes.
• Collaborative Debugging: Working with an agent to investigate a rendering issue or a bug while keeping the entire diagnostic conversation linked to the code changes.
• Agent-Assisted Exploration: Exploring complex architectural changes or cache invalidation logic by tasking an agent with experiments and comparing the results within a shared thread.