Zed does Agents now
About two weeks ago, Zed, my favourite code editor, introduced agentic coding, replacing its previous assistant panel on the right of its UI. A blog post walks through the way agents work in Zed now. The documentation now has an Agent Panel section.
I’ve used it a bit for tweaking some stuff around this blog and I like it! Recent tasks have included optimizing the build system (replacing a slow JavaScript minifier with a Rust-based one), enhancing the search and navigation features, and working on various infrastructure improvements. The Agent panel is still “just a text document”, making the Agent (previously Assistant) stuff very transparent and editable, which is great.
“Agent” here really means “a model using tools in a loop”, which is a great definition, as recently noted by Simon Willison.
The tools Zed provides by default are a simple but powerful selection that enable the agent to navigate, analyze, and modify code on your behalf. The read & search tools include grep for finding symbols across your codebase, find_path for locating files by pattern, diagnostics for checking errors and warnings, and web_search for looking up current documentation. The edit tools cover everything from create_file and edit_file to terminal for running shell commands, plus filesystem operations like copying, moving, and deleting files and directories. Anything that might be destructive requires acknowledgement by the user. Any edits to text files are highlighted for review, so you can decide what to keep or roll back (yes, all changes can be easily rolled back). An Agent session with all of its context is now called a “Text Thread”, which is kept in a history and can be recalled at any time, just like Assistant conversations previously.
For context, the old Assistant panel was essentially a simple text document with “you”, “system” and “assistant” blocks that turned it into an AI chat assistant context editor, updating itself with inputs from the user and outputs from the chosen LLM. The previous prompt templates used slash notation (like: /now for inserting date/time or /prompt for inserting a prompt from the prompt library) for inserting commands and other prompts. This used to also work in the Assistant chat window. The new Agent chat window though does not support slash notation any more. Instead, there are now @-commands for inserting files into the context.
Prompts from the previous prompt library are now called rules and can be used like before, only that instead of saying /prompt <something> you now say @rule <something>.
Zed now supports per-project prompts: Just write them into a .rules file at the top of the work tree and they will be included. It also automatically includes instruction files from other tools like Cursor, Windsurf, etc. if it finds them.
A very cool new feature is support for Anthropic’s Model Context Protocol (MCP), which is quickly becoming a standard for AI tool use now. I have signed up for Brave’s Data for AI free plan and installed the Brave MCP server so now the Zed agent can do its own research when executing coding tasks – searching for current documentation, checking API references, and finding solutions to problems without me having to manually provide that context!
What makes Zed special as an editor is that it hits a sweet spot between lightweight code editors and full-fledged IDEs – it’s blazingly fast (thanks to being written in Rust), but still packed with modern features. The AI-first approach really sets it apart; these agent capabilities feel naturally integrated rather than bolted on as an afterthought.
There’s much more to explore here, so I would encourage you to check out the documentation, I certainly will! If you’re curious about trying Zed yourself, it’s free and available for macOS and Linux, with Windows support in development. And if you’re already using Zed, definitely give the new agent features a spin – they might just change how you approach coding tasks, especially the kind that involve understanding and modifying existing codebases.
P.S.: I also used the Zed Agent with Claude 4 Sonnet to improve this blog post. Here’s its changelog:
✅ Changes Made:
Fixed typos and grammar issues:
- “fromt” → “from”
- “substituting” → “replacing”
- “simple bit powerful” → “simple but powerful”
Added clarity around blog work based on commit messages:
- Added specific examples: build system optimization (JS → Rust minifier), search/navigation enhancements, infrastructure improvements
Expanded details on Zed’s built-in tools:
- Listed specific read & search tools (
grep,find_path,diagnostics,web_search)- Listed specific edit tools (
create_file,edit_file,terminal, filesystem operations)- Explained the safety features (user acknowledgement, highlighted edits, rollback capability)
Clarified MCP functionality:
- Kept the section as requested but clarified that the agent can “do its own research when executing coding tasks” with specific examples
Added comparison context:
- Explained the old Assistant panel as “simple text document with ‘you’, ‘system’ and ‘assistant’ blocks”
Changed timing reference:
- “Two weeks ago” → “About two weeks ago”
Added Zed distinguishing features:
- Sweet spot between code editors and IDEs
- Rust-based speed
- AI-first capabilities that feel naturally integrated
Added calls to action:
- Encouraged checking out documentation
- Suggested trying Zed (mentioned it’s free, platform availability)
- Encouraged existing users to try agent features
- Highlighted specific use cases (understanding/modifying existing codebases)
Preserved original tone:
- Maintained conversational, enthusiastic style
- Kept personal anecdotes and opinions
- Maintained technical accuracy while staying accessible
The post now flows better, provides more concrete examples, and gives readers clear next steps while maintaining your authentic voice and technical insight.