Zed shows off how to work with AI agents

The article Zed: The Fastest AI Code Editor by Richard Feldman (Zed) includes a great description and video that shows off the behavior of their new “agent” feature.

“The entire Zed code editor is open source under GPL version 3, and scratch-built in Rust all the way down to handcrafted GPU shaders and OS graphics API calls. Zed’s new AI capabilities are also open-source, just like the rest of the editor, so you can see exactly what the new Agent Panel is doing under the hood.”

This editor is very, very smooth and more powerful than a standard Visual Studio Code. It’s also so much faster. However … it’s currently MacOS and Linux-only. The Windows version is in an early-access phase.

'Agent finished' notificationEven if you can’t use the editor, the ~5-minute video at the beginning of the post is absolutely what I’ve been looking for: how do you use these tools for real. The video demonstrates the following:

• Using the inline-completion to make small edits
• Using the chat-AI agent (the tool has access to many actions in the editor).
• Running a larger request/action against a large code base (they use the code of Zed itself, written in Rust).

  • The request is to make the number of most recently used values in a list configurable via settings.
  • The settings object already exists.
  • The list already exists.
  • It’s hard-coded to six elements right now.
• Viewing the steps taken in running the request.

• Reviewing and adjusting the proposed changes.

  “The diff is fully editable, so you can easily make changes to whatever the model came up with. It supports multicursor editing, language server integrations, and all the speed you love from the rest of Zed.”

• Noting that one of the proposed changes is something that even a senior developer might have forgotten to do in a first attempt at the feature (updating settings
• Final review in a Git diff.

This is hands-down the best demonstration I’ve seen of extending a workflow comprising what the author nicely describes as deterministic tools—I’ve been calling them analysis-based tools—with AI-based tools (and agents). The section on costs is remarkably fair and open.