Start building with Zed's agentic editing by accessing the Agent Panel, either via the sparkles icon button in the status bar (✨, bottom right), or by running assistant: toggle focus in the command palette.
Models
During the beta launch, agentic editing is only available for the Claude family of models from Anthropic. Tool calling support will be added for other models in the coming days!
Similar to our current AI assistant offering, we provide both a hosted service ("Zed Pro") to access Claude models where we bill you for your usage, or allow you to bring your own API key and be billed directly through Anthropic. Visit the docs for further configuration instructions on how to bring your own API keys.
Please note that we may shift our billing and pricing as we learn more about usage. We are not looking to profiteer from AI models, but LLMs are expensive to run, and we need to cover our costs.
Profiles
The Agent Panel comes with two out-of-the-box profiles: "Write", optimized to edit and change files in your codebase, and "Ask", better suited for research and general conversation.
You can customize the included profiles as well as create new ones from scratch. Profiles are, for now, a combination of tools and Context Servers that you can arrange as desired.
Checkpoints
Every time that the agent edits something in your codebase, Zed creates a checkpoint that allows you to restore your code, and the conversation, to a previous state.
Reviewing Changes
You can chose to keep or discard the edits performed by the agent in a dedicated multibuffer diff view. And, as soon as you're happy with it all, you can use the Git Panel to commit them.
Extensibile
Enhance the Agent Panel capabilities by installing Context Servers (via the Model Context Protocol), which can be done either via Zed's extension store (see the ones available already) or directly via the Agent Settings view by clicking on the "Add Context Server" button.
Explicit Context
Similar to slash commands in our original assistant panel, you can add specific context by @-mentioning files, directories, or previous threads in the Agent Panel textarea. These are useful if you know exactly what you want the Agent to look at, which potentially speeds up and improves the quality of its response.
Support for adding content from your terminal, diagnostics, and more as context is coming in the near future.
Feedback
We’d love to hear how your experience with the beta goes. If you run into any bugs, let us know! To test the beta, you’ll need to be on the latest preview build.
After downloading and installing the latest preview, you will need to sign in via the GitHub auth flow, by clicking the "sign in" button in the upper right hand corner of Zed.
You can send us feedback by joining the “agent-beta” Discord channel, replying to this email, or rating the Agent threads directly via Zed.