I personally explicitly avoid parallel agents, since it creates too much cognitive debt, and sometimes an agent may need steering towards an architectally sane solution mid-work.
I'm buying into this workflow more the more I use it, but the real gamechanger is (a) parallel threads in worktrees, with (b) enough lifecycle hooks to treat them similarly to spinning up a VM.
Specifically for me that means that after I create a worktree I get some local config files copied over and Postgres duplicating my local dev and test databases so I can test in isolation, and then when I close out a worktree it deletes those databases.
The best at that that I've found is Conductor, but I can't use it at work because we only have Copilot and they're locked to a Claude/Codex backend. Arbor is close, but it's not under as active development and has a lot of rough edges. Opencode GUI has create hooks but not teardown.
If Zed can hook that up _and_ also keep its great editor roots, that'll definitely be a game changer.
This is helpful to know - we're working on adding more agents, Copilot and OpenCode harnesses are among the most popular requests.
We also recently built an escape hatch. If you turn on Settings → Experimental → Big Terminal Mode you can create new terminals in the center panel (with ⌘⇧T) and use any agent you'd like (Copilot, OpenCode, etc). It isn't the best experience because you don't get notifications etc (yet), but at least it lets you use the harness you'd like until we build out the first-class UI for it.
Send me feedback anytime, I'm charlie@conductor.build.
I’m interested in these types of tools, but the biggest issue I am currently having in my workflow is manually verifying everything is working. Tests and stuff are nice, but typically if there is a bug, the AI agents enshrine it in th tests.
However with worktrees I am not really able to easily copy secrets, etc to run my app, ports conflict, I end up with a bunch of separate dbs and services, etc.
Does conductor help with this? Have you all found any useful ways of making this easier or more automated?
I was an early user of conductor and used it a lot (like from maybe oct to Jan). But then there was some bug where maybe it wouldn’t release file descriptors or something where my laptop needed to be rebooted twice a day. So I stopped using it months ago.
But I’ve tried to reinstall it since and it just gets stuck in a weird infinite loop.
I liked conductor though. Hope you are able to fix those bugs and I can try again in a few weeks.
Howdy, love your product. My coworker and I have literally been begging our coworkers to give it a try. To me, it was the key tool to 'get it' (coming from Cursor) when going from single threaded agentic dev to parallel agentic dev. Please keep up the great work.
In VSCode, I use https://github.com/jackiotyu/git-worktree-manager for the same purpose - the extension has before create/before destroy WT hooks which you can run anything from. Mine symlinks workspace file from main checkout, installs packages and copies over some files. Very handy.
You should check out Ouijit [1] - I use it regularly for work and it's nice because it focuses on the environment that you want, and just gives you a shell that you can use any tooling in, as well as VM isolation per worktree if needed.
No need for any AI-specific tool, this is exactly what devcontainer is for! Just tell your agent to use devcontainer up (and docker compose down the other way).
Devcontainers always disappointed me. The sales pitch is that everyone uses the same container, but that's not accurate. Everyone builds a container from the same config and it'll be similar, but it takes a ton of effort to make sure it's identical.
The idea that a devcontainer gets built on-demand instead of checked out like 'docker pull ..." has always felt weird to me. It's so close to being awesome, but ends up being barely useful.
Or maybe I'm wrong. Is there a way to checkout an immutable devcontainer?
tried parallel agents for a sprint and bounced off it. the worktree dance is fine, real blocker for us was test data isolation. scoped postgres schemas per branch worked, but reasoning about which agent broke teh shared migration when three of them touch it got old fast. we just run one agent at a time now and go for a walk.
It's pretty clear by this point that everyone is going towards parallel agents and worktrees, but TBH I am surprised to see an offering from Zed, seeing how heavy they lean into being editor-heavy and having AI features be strictly optional.
The key advantages Zed has are being agent-agnostic (so not a first party UI like Claude/Codex/Cursor Desktop), supporting multiple repositories on the same agent via creating a worktree for each automatically, and having a high quality custom agent UI rather than wrapping over CLIs (I've used their IDE's agent UI in the past and it's great). AFAIK, this is the first mainstream tool that supports all of these features.
yeah, but they don't support a lot of the features Claude does like MCP integration. we've booked this up to logfire (telemetry provider) and it's a game changer when doing optimisation or diagnosis of bugs. plugins and skills also missing. but it is nice to switch between provider easily
The new default layout is exactly backwards of what I want.
It should go: project tree | text editor | agent view | threads
Not to mention on most laptops you'll only have room for about two panes at a time. So they should be focusing on pane management and making it easy to swap between views. Not highlighting 4 pane workflows. Unless you have an ultra-wide monitor, I'd rather have a separate Agents window.
I use Zed a lot and this is a minor (can be configured) but telling design decision that really bothers me. How long until they decide editing itself isn't worth it anymore? Stop supporting VI mode?
Yeah, first thing I did was change all the locations back. I really dislike the automatic layout change they pushed.
Also, watching the change notes, most effort seems to be focused on agents this days, which is a bit worrying. I love Zed because it's great editor that also knows a bit about agents; I don't want it to continue pivoting towards pushing agent management deeper and deeper into the experience.
Personally, I still pay Zed because I think they’re doing great foundational work. However I’ve now almost entirely migrated to helix for editing needs.
> I don't want it to continue pivoting towards pushing agent management deeper and deeper into the experience.
AI and agent tooling is the only way this VC-backed company is going to make any money. Nobody is going to pay for a text editor, so they've had no choice but to enshitify it by chasing AI features for the last year+.
I tried Zed and was really convinced that I could use it full-time, but the lack of extensions (TODO highlight, TabOut) + other QoL (line number goto as easy as VSCode + tab filter like another comment said).
I also thought it was odd that I can't configure the font size in the git commit message editor.
On recent additions, the dev container integration was great.
The Parallel agents feature seems to be built around git worktrees or local projects. But I think local project mode defeats the point. In my day to day development workflow I've switched fully to jj and jj workspaces, so this isn't something I'll be using until Zed gets jj support. Also this change ended up with a reshuffled layout I didn't really expect and I can't really figure out how to revert it now
I find parallel agents to be an exception rather than a norm. Maybe I’m the problem?
For those exceptional cases, opening a few more terminals gets the job done. It’s unclear to me if this needs to be the primary workflow. My brain naturally does better on deep work on one problem..
I am exactly the same, except I am really excited about this update! It’s not so much “in parallel” but being able to easily jump between threads. It allows me to dive into misc investigations in a side thread without derailing some main context where I’m doing the main editing.
I have a coding agent https://github.com/hsaliak/std_slop where the sessions are in SQL ledger. So /session [new, clone, deletes, undo] are supported and all sessions are persistent. Cloning lets you 'fork' the context and undo lets you roll back, basically solving the problem you state above.
Sessions are linear though, so you cant do this _while_ an existing session is cooking.
That said, I am excited about this update too, I've been playing with ACP support and Zed's UX was bare bones. I want to run my agent with multiple workers now, and see what happens.
I have historically not used them but I want to start so that some of the spin up / tear down work of doing any particular task can happen in isolation. For example, drafting a change before I start editing, checking out and setting up code from a branch before I do a review, etc.
Yeah, we are getting to the point where you having a 4K monitor for your editor becomes a requirement rather than a "nice to have". Until now you have the agent, the editor and files/git/whatever, if you add a fourth panel to that, it becomes pretty cramped at lower resolutions. Fortunately, I do have a 4K monitor, but until now I used to have the editor on one half of it and another window (a browser most of the time) on the other half. So having to use the editor full screen is still slightly annoying.
Of course, that's only the default layout. I'm not familiar enough with Zed, but there's probably a way to change it? In JetBrains IDEs, you can configure panels to sit at the top left/bottom left/left bottom/right bottom/bottom right/top right side and show/hide them with one click (if only one panel on the respective side is shown, it will take up the full space). So you could have files at the top left and the agent panel at the bottom left. And the code editor is of course still the "centerpiece" in the middle.
Luckily it's very easy to change, although a bit unintuitive for new users. You right-click the small icon for each type of panel in the bottom bar, and select where you want it to be docked. Left click toggles the panel into view.
It will also pull more users in. I don't want to be looking at code. I want more of a codex style app where it's easy to shove all my projects in one place and context switch endlessly.
This was also my first impression. But it seems to me the changes are mostly about swapping what panels dock where (left or right) and maybe some additions/tweaks around the AI panels. On macOS these are still the same:
⌘B : toggle the left dock
⌘R : toggle the right dock
If you opt-in to the new layout, the panels that used to sit in the left dock are now in the right dock. I will give it a try even for classic coding. One can change what panels get docked where from the settings window.
Warp launched something similar a week or so ago, but the Zed implementation I find a lot more logical. Will give Zed another try, as I’m overdue for my monthly “maybe I should try this terminal/IDE” itch.
I like Warp but something about it is very opaque and confusing. Maybe it has a learning curve I haven't committed to, or it's just very alpha and evolving often.
Agreed, I exclusively use Warp for server maintenance and ssh'ing into servers, it does that better than Claude itself but the UI is always confusing, especially after their recent changes.
I'm having a hard time adjusting to the Project Panel on the right (and, at least for me, hidden by default) - seems like they're trying to bury the concept of a 'file'?
It's certainly interesting though, and I'll give it some time - the post says "It feels more natural once you've spent a little time with it"
I think it's more that you don't need to look at the actual tree structure all that often, especially when most codebases follow logical and conventionalised structures you remember these structures and can use Command-P to search for files individually to open and the tree structure becomes something you peek at infrequently to ensure that everything is going in it's right place.
I disagree. All their "AI slopshit" features are useful, and the fact that you can just.. turn them off and get a great editor regardless is what I like about it.
I would love to unleash parallel agents, but I am still checking every single edit while enforcing minimal, stateless, modular code, and I have the AI check in with me before writing the next file.
A lot of times, I find it has incredibly stupid ideas and tends to make the code very messy. I would love to figure out how to stop that from happening automatically.
The upside of checking in on the code, though, is that I can come up with smart directions for the AI from both a product and tech perspective. This is especially helpful when the dumb suggestions add a lot of complexity.
I think it's like when a product person asks for a new feature, or when a founder building their own product selects which feature is smarter to build and how.
I'm expecting we'll likely end back up on agents making PRs, and having to review them. Either that or giving up on quality etc/dealing with very messy code. I've been trying various automated testing/linting/etc strategies, and they only work so well.
That would be a nightmare. One thing is to review a PR generated by a human using AI and caring about the code; another is reviewing wild agents, especially when they make changes everywhere
I'm not excited about it, but the only main way I've been able to discover LLM-isms that sneak in are
1. via seeing them glimpse by in the agents' window as its making edits (e.g. manual oversight), or
2. when running into an unexpected issue down the line.
If LLMs cannot automatically generate high quality code, it seems like it may be difficult to automatically notice when they generate bad code.
I think the issue is deeper than prompts, agents.md, smart flows, etc. I think the problem is that LLMs are searchers, trained on preferring some results. So, if the dumb solution is there, and the smart solution is not there, they won't spit it out.
I really want to like Zed, but for some reason the way it interfaces with the TypeScript language server is dog slow compared to VS Code and its derivatives.
Check the lsp servers that run when you are editing TS file. I noticed something like 5-6 when editing a TSX file. There was even tailwind thrown in for a good measure (the project has no tailwind). All that buttery smooth 120fps wasted on poor resource management when it comes to LSPs.
Same. Zed is so fast it's shocking. But it was using 2-3x as much RAM as VSCode or Cursor with TypeScript and the language server crashed a lot. Given that work is a TypeScript monorepo, that was a dealbreaker for me.
What are parallel agents worth to professional engineers if reviewing the code is a pain (aka non existent) in Zed? Please add proper code review tools (compare with branch|file|revision) and GH pr review tools like IntelliJ!
Zed is probably the best text editor in the last 10 years. It has quirks, but it is insanely powerful and capable out of the box. I don't even bother trying to setup Neovim because of Zed. They let people for PRs for missing vim features for their vim emulation, and its insanely capable.
I hope someday they get the funding they deserve, because it has insane potential. It's why I subscribe to their pay plan, even if I dont use it all the time, I want them to succeed.
I do wish they'd focus on closing the gap to Jetbrains by implementing the QOL features that are missing. I understand they have to do what VC wants to see, but this agentic stuff is so tiring.
I give them a try about twice a year. I write a lot of Rust which should be squarely in their wheelhouse.
This last time I was pleasantly surprised to find they mostly fixed their SSH remote editing support. But then it started truncating rustc inline error messages and I couldn’t figure out how to view the whole thing easily. When you’re just trying to get something done little bits like this can add up quickly. Punted back to Cursor for now.
I don't like the way remote editing works with plugins. IIRC, the remote agent pulls the plugins from the connecting client. I get why it's done like that, but I'd way rather have it go the opposite direction.
I want a setup where I can have an immutable devcontainer with local copies of everything I need to develop 100% offline; dependencies, tools, etc.. Having my local editor pull plugins from a devcontainer for the project seems to make more sense to me.
I didn't dig in too much. Maybe there's a way to make it work somehow.
I give them props for writing a truly responsive editor that's easy to use, and can switch to keybinding sets from other editors. (So you don't have to learn new ones) And I find it works more naturally for multi-file projects than VsCode, and Sublime.
In terms of in-line instantaneous error highlighting, introspection, refactoring, and autocomplete, it's not on the same level as JetBrains.
Love Zed for my personal projects at home with Openrouter. But I cant use it at work as I'm forced to use claude code at work and find the VSCode+claude code combo better. I know zed has claudecode integration, but I found it to be very weak compared to zedagent+claudemodel. Has this improved?
Do you mind sharing what your spend is like on Openrouter? I pay for the $200 Claude Code plan just so I never have to think about usage or feel like I'm paying per prompt. But I nearly max out my usage each week and I have no idea what that would cost me on Openrouter. I hear a lot of people are using MiniMax of Openrouter. Is that what you do?
What I want is a stateful file-writing layer that is aware of all clients (aka agents and humans) and their activity. It provides its own locking mechanisms, and prevents agents from overwriting each others work. That way you could have multiple agents operating on the same codebase, without having to futz with worktrees and all that.
tried using zed but when i use codex cli, when agent updates file, it doesn't even register in the "buffer" (why call it buffer? trying to be different?), i have to close the file and open it back up again for it to update the file contents which creates bugs where i edit some other thing in file and save (no "overwrite warning" which i get in vscode) it just overwrites the file....so the editor is just plain buggy.
I loved zed for over 1 year, told for everybody to use it, because it was so fast and great.
But now using claude-code,gemini-cli,codex,etc it just seems less relevant. Just opened nvim with lazyvim and it feels nice, since I'm in terminal anyway it just feels more natural.
Still have zed opened, still like it but I guess honeymoon is over.
I really want to like zed but it claims all over the place to be be built with native windows code but it's the only OS I can't use the OS title bar and this weird menubar that opens on hover?
The claim really falls flat the moment you open it up and it is "native but not as you know it" providing the same UX experience of any other electron junk.
I've used Zed since the very beginning and I remain a fan. Its LLM integration so far has been a lot more pleasant than what I see in others and the editor is perfectly usable without using LLMs.
Its multi buffer and speed sound trivial but using anything else feels wrong now.
I liked the idea of the new layout with the agent thread on the left, it goes hand-in-hand with having multiple threads that are easy to switch between and running concurrently, but I switched back because my file tree disappeared and I couldn't easily see how to add it back
In the new layout, the project panel and git panel are just moved to the right side, so that the agent panel could be on the left, and you could still view both at the same time.
I wonder if you could use something like vLLM and have these subagents max out your local gpu. I’ve been looking for using a local model because I’m tired of rate limits of the cloud and also would really like to make use of my local gpu when I’m working (5090h even if it’s not on my computer I’m typing on.
Love your editor again
Zed is a minimal code editor crafted for speed and collaboration with humans and AI.
At home, I don't use any AI when coding, to keep my brain sharp. But it's clear that Zed's focus is on AI integration because that's where the money's going (seriously, where is the setting to have a different ui icon size vs ui font size). Is there any editor still being being developed and focusing on the experience of coding by hand?
> seriously, where is the setting to have a different ui size vs ui font size
Search for font size in preferences.
You'll see a 'font size' under 'buffer' (editor), under 'UI Font', and under 'Agent Panel' to let you control font sizes in all of those places independently.
> Is there any editor still being being developed and focusing on the experience of coding by hand?
Zed lets you hand-edit too! It's fast and decent. vim, neovim, Emacs, Helix, and JetBrains products continue to do that well too. There are still more traditional IDEs/editors than pure AI ones.
You can also toggle AI features off in Zed from preferences if you want to.
ah I forgot a word, I meant the ui icon size. If I bump up the ui font size so that I can distinguish the icons apart on my large monitor, the ui text becomes comically large.
I do use Zed without AI features, it's just a bit of a disappointment (though understandable) since it was originally marketed as just a nice speedy editor.
I don't think changing icon size independent of UI font size would be a dealbreaker for many. (I'm quite happy having icons that scale in line with font size, but then I use the Material Icon Theme, which is easy to scan at most sizes.)
Just injecting this here: What I've been missing is an equivalent for GitHub's "blame prior revision" feature to quickly follow through the history of individual source lines.
I use zed every day, but the git view is still buggy and unreliable for me. It often chooses the wrong folder---a subfolder rather than the root of my git project---and shows that there are no changes. And sometimes, even when the right folder is selected, it will still incorrectly say there are no changes.
Another bug is that clicking on new files in the git UI doesn't bring up their diff---you need to click them twice to open the file. I'd like to have those files included in the diff view just like other files are.
I'd love for there to be an easier way to get to the "view file history" view---I didn't realize that it existed until I tried searching for it. I'd love a "line history" view like GitLens has, as well.
If you're taking suggestions I'd like to be able to see when I'm over the 72 character limit, last time I checked there was no way to know inside Zed when writing the commit message though I might be wrong. Other than that I think Zed's great and multi-buffer editing is really swell.
I'm building Paneditor with that focus. My goal is to close the gap so that (competent) humans can work with code at levels of editing throughput usually reserved for users of LLMs.
I like LLMs, I like Zed, but I turn off the AI features. I rather have Claude or Open Code in a container with only access to a mounted folder, or use a local model.
And Zed lets me do that while remaining fast and minimal.
As for (even more) minimal editors, perhaps just Gnome Edit? Or Kate?
ecode [0] is ridiculously fast -- makes Zed feel like molasses -- and quite customizable. It's still early in development and mostly just made by one guy (who is also developing the GUI framework used), so progress is slow and there are some rough edges, but it has all the important stuff and quite a few niceties too. Really cool project, reminds me of Sublime.
The thing is, they have to monetize somehow. There's a setting to turn all AI features off with one toggle and you're back to an 'editor still being being developed and focusing on the experience of coding by hand'
100%. I recently got rid of my lsp-booster and similar kludges because the builtin language server client (eglot) is now fast enough without it, even on large projects.
And if you want AI integration at your choice and control, agent-shell (and chatgpt-shell, which is LLM-agnostic despite the name) are great packages. They’re totally hackable with elisp like you’d expect, which I personally haven’t done a ton with, because I use AI pretty sparingly, but I imagine the crowd here could come up with plenty of ideas for how to program your editor and your agent interface together.
> You do realize software engineering is in the midst of a tectonic shift?
As an everyday user of AI, both at work and privately, I am not that convinced. The biggest effect I've seen so far is demand for faster work because "everything is faster with agents", but software quality is slowly dropping in software I see around me.
Current AI is very useful as a trivia engine and as a language manipulation tool - i.e. it can quickly extract information from a huge amount of text. But it still sucks when writing new things.
Admittedly, here has been much progress, but it seems to be slowing down. Money is drying out, models are getting nerfed, and only better scaffolding and workflows are making it better. Unless they build 100x more data centers, I don't see models getting significantly better.
What will convince people is what they see with their own eyes. Not yet another proclamation that the revolution is happening right now.
Like this.[1]
> AI-assisted coding has become the norm and with tools like Cursor, GitHub Copilot, Claude Code, Codex, we are increasingly letting models touch our code.
... how is it good literaly style to both (1) assume that something is the norm, and (2) use a long intro-sentence to state that something is the norm? Pick a lane—either it is the norm and you don’t need to state it or it isn’t and you need to set the stage. Stating the apparently obvious makes your (their) writing read like a eighth grade paper.
In short I’m agnostic as far as proclamations go. ;)
I like Zed but the defaults for tabs (CTRL+Tab and CTRL+Shift+Tab step you through some diabolical tab selector in what is allegedly (but not really) recency order instead of just left/right. It really made me question their sanity.
Is this any different from a setup where I use a terminal with tabs and splits, running my favorite editor in one or more panes, and several agents (Claude Code and Codex) in several other panes and tabs?
Edit: Although I can integrate an agent in NeoVim, I don’t do it. I want to use my editor solely for that purpose, while the rest (versioning, agentic coding, git client, etc.) is done in the terminal. My NeoVim setup is simple and fast, which is why I prefer it over any other IDE or editor. Especially with the native package manager in the latest version. I also replaced BBEdit by installing Neovide, a GUI version of NeoVim. It starts in a split second and is incredibly smooth and fast. And it’s so enjoyable to work with that I use it as my preferred frontend to Obsidian.
The worktree part is the easy half. Running parallel Claude Code subagents, the bottleneck is never "can they run without stomping each other's files." That's solved the moment each one has its own checkout.
The hard problem is architectural consistency. Agent A renames a type to X. Agent B, in a different worktree, independently renames the same type to Y because neither saw the other's decision. When you merge, neither worktree is "wrong" but the code is incoherent. You need either a shared decision log that every agent reads before starting, or an orchestrator that hands out scope narrow enough that no two agents can collide.
Zed's post is solving filesystem-level parallelism. The harder coordination problem is semantic, and that's where time savings from parallelization go to die.
I personally explicitly avoid parallel agents, since it creates too much cognitive debt, and sometimes an agent may need steering towards an architectally sane solution mid-work.
Same here. Reviewing gets harder too and multi tasking kills any kind of productivity if you need to review the code then.
My approach these days is to do one change at a time, until I can fully merge it with confidence.
I'm buying into this workflow more the more I use it, but the real gamechanger is (a) parallel threads in worktrees, with (b) enough lifecycle hooks to treat them similarly to spinning up a VM.
Specifically for me that means that after I create a worktree I get some local config files copied over and Postgres duplicating my local dev and test databases so I can test in isolation, and then when I close out a worktree it deletes those databases.
The best at that that I've found is Conductor, but I can't use it at work because we only have Copilot and they're locked to a Claude/Codex backend. Arbor is close, but it's not under as active development and has a lot of rough edges. Opencode GUI has create hooks but not teardown.
If Zed can hook that up _and_ also keep its great editor roots, that'll definitely be a game changer.
I've built multi-agent workflow on JJ workspaces - independent from specific agent. You can run Codex, Claude or anything else - https://www.visualjj.com/learn/parallel-ai-agents
Love to hear it! (Conductor founder here).
This is helpful to know - we're working on adding more agents, Copilot and OpenCode harnesses are among the most popular requests.
We also recently built an escape hatch. If you turn on Settings → Experimental → Big Terminal Mode you can create new terminals in the center panel (with ⌘⇧T) and use any agent you'd like (Copilot, OpenCode, etc). It isn't the best experience because you don't get notifications etc (yet), but at least it lets you use the harness you'd like until we build out the first-class UI for it.
Send me feedback anytime, I'm charlie@conductor.build.
Anyone know of a similar tool to conductor for Linux?
I’m interested in these types of tools, but the biggest issue I am currently having in my workflow is manually verifying everything is working. Tests and stuff are nice, but typically if there is a bug, the AI agents enshrine it in th tests.
However with worktrees I am not really able to easily copy secrets, etc to run my app, ports conflict, I end up with a bunch of separate dbs and services, etc.
Does conductor help with this? Have you all found any useful ways of making this easier or more automated?
Conductor has a feature that lets you define a script to copy .env and other things as needed from main repo into new worktrees on creation.
I don't use it much anymore, but last I did I would run into port conflicts with docker projects.
I was an early user of conductor and used it a lot (like from maybe oct to Jan). But then there was some bug where maybe it wouldn’t release file descriptors or something where my laptop needed to be rebooted twice a day. So I stopped using it months ago.
But I’ve tried to reinstall it since and it just gets stuck in a weird infinite loop.
I liked conductor though. Hope you are able to fix those bugs and I can try again in a few weeks.
Howdy, love your product. My coworker and I have literally been begging our coworkers to give it a try. To me, it was the key tool to 'get it' (coming from Cursor) when going from single threaded agentic dev to parallel agentic dev. Please keep up the great work.
In VSCode, I use https://github.com/jackiotyu/git-worktree-manager for the same purpose - the extension has before create/before destroy WT hooks which you can run anything from. Mine symlinks workspace file from main checkout, installs packages and copies over some files. Very handy.
You should check out Ouijit [1] - I use it regularly for work and it's nice because it focuses on the environment that you want, and just gives you a shell that you can use any tooling in, as well as VM isolation per worktree if needed.
[1]: https://github.com/ouijit/ouijit
No need for any AI-specific tool, this is exactly what devcontainer is for! Just tell your agent to use devcontainer up (and docker compose down the other way).
Devcontainers always disappointed me. The sales pitch is that everyone uses the same container, but that's not accurate. Everyone builds a container from the same config and it'll be similar, but it takes a ton of effort to make sure it's identical.
The idea that a devcontainer gets built on-demand instead of checked out like 'docker pull ..." has always felt weird to me. It's so close to being awesome, but ends up being barely useful.
Or maybe I'm wrong. Is there a way to checkout an immutable devcontainer?
Just specify “image” https://containers.dev/implementors/json_reference/
> The best at that that I've found is Conductor, but I can't use it at work because we only have Copilot and they're locked to a Claude/Codex backend.
FYI, you can use Copilot directly in Zed!
We support create git worktree hooks too
Is there an example of this?
The worktree create hook is documented https://zed.dev/docs/tasks?highlight=hook#hooks
And then I guess setting up tasks for the cleanup part, but it'd be great to see that get automated too so I don't need to remember it.
self-promotion, but check out worktrunk.dev
by far the most popular worktree manager
tried parallel agents for a sprint and bounced off it. the worktree dance is fine, real blocker for us was test data isolation. scoped postgres schemas per branch worked, but reasoning about which agent broke teh shared migration when three of them touch it got old fast. we just run one agent at a time now and go for a walk.
It's pretty clear by this point that everyone is going towards parallel agents and worktrees, but TBH I am surprised to see an offering from Zed, seeing how heavy they lean into being editor-heavy and having AI features be strictly optional.
The key advantages Zed has are being agent-agnostic (so not a first party UI like Claude/Codex/Cursor Desktop), supporting multiple repositories on the same agent via creating a worktree for each automatically, and having a high quality custom agent UI rather than wrapping over CLIs (I've used their IDE's agent UI in the past and it's great). AFAIK, this is the first mainstream tool that supports all of these features.
yeah, but they don't support a lot of the features Claude does like MCP integration. we've booked this up to logfire (telemetry provider) and it's a game changer when doing optimisation or diagnosis of bugs. plugins and skills also missing. but it is nice to switch between provider easily
Little historic sidenote: Zed was the first MCP client next to Claude Desktop
The built in agent supports MCPs, as do the agents you can use through ACP that do already.
This is false. Zed does support skills and also MCP servers.
I’m using MCPs with Claude Code in the Zed UI
The new default layout is exactly backwards of what I want.
It should go: project tree | text editor | agent view | threads
Not to mention on most laptops you'll only have room for about two panes at a time. So they should be focusing on pane management and making it easy to swap between views. Not highlighting 4 pane workflows. Unless you have an ultra-wide monitor, I'd rather have a separate Agents window.
I use Zed a lot and this is a minor (can be configured) but telling design decision that really bothers me. How long until they decide editing itself isn't worth it anymore? Stop supporting VI mode?
Yeah, first thing I did was change all the locations back. I really dislike the automatic layout change they pushed.
Also, watching the change notes, most effort seems to be focused on agents this days, which is a bit worrying. I love Zed because it's great editor that also knows a bit about agents; I don't want it to continue pivoting towards pushing agent management deeper and deeper into the experience.
$VC disagrees w/ you on this.
Unfortunately.
Personally, I still pay Zed because I think they’re doing great foundational work. However I’ve now almost entirely migrated to helix for editing needs.
> I don't want it to continue pivoting towards pushing agent management deeper and deeper into the experience.
AI and agent tooling is the only way this VC-backed company is going to make any money. Nobody is going to pay for a text editor, so they've had no choice but to enshitify it by chasing AI features for the last year+.
I tried Zed and was really convinced that I could use it full-time, but the lack of extensions (TODO highlight, TabOut) + other QoL (line number goto as easy as VSCode + tab filter like another comment said).
I also thought it was odd that I can't configure the font size in the git commit message editor.
On recent additions, the dev container integration was great.
Rooting for you Zed!
use zed agents to build the zed extensions you miss
FYI there is now an extension for TODO highlighting. Not at my machine but it’s something like “comments highlighter”.
https://github.com/thedadams/zed-comment
The Parallel agents feature seems to be built around git worktrees or local projects. But I think local project mode defeats the point. In my day to day development workflow I've switched fully to jj and jj workspaces, so this isn't something I'll be using until Zed gets jj support. Also this change ended up with a reshuffled layout I didn't really expect and I can't really figure out how to revert it now
I find parallel agents to be an exception rather than a norm. Maybe I’m the problem? For those exceptional cases, opening a few more terminals gets the job done. It’s unclear to me if this needs to be the primary workflow. My brain naturally does better on deep work on one problem..
I am exactly the same, except I am really excited about this update! It’s not so much “in parallel” but being able to easily jump between threads. It allows me to dive into misc investigations in a side thread without derailing some main context where I’m doing the main editing.
I have a coding agent https://github.com/hsaliak/std_slop where the sessions are in SQL ledger. So /session [new, clone, deletes, undo] are supported and all sessions are persistent. Cloning lets you 'fork' the context and undo lets you roll back, basically solving the problem you state above.
Sessions are linear though, so you cant do this _while_ an existing session is cooking.
That said, I am excited about this update too, I've been playing with ACP support and Zed's UX was bare bones. I want to run my agent with multiple workers now, and see what happens.
I also have this workflow, it’s like you’re on the side quests or on the main story; those are not necessarily in parallel
I have historically not used them but I want to start so that some of the spin up / tear down work of doing any particular task can happen in isolation. For example, drafting a change before I start editing, checking out and setting up code from a branch before I do a review, etc.
I personally don't love the idea of the default layout pushing aside my code and filetree to make space for AI tools
I really like Zed, I use it every day. But, if I'd seen this layout when I first installed, I never would have taken it seriously
I imagine this will push some new users away
> I imagine this will push some new users away
I suspect it will gain them more users than it will lose
Most other tools doing this are heavy, buggy, and built on electron
Yeah, we are getting to the point where you having a 4K monitor for your editor becomes a requirement rather than a "nice to have". Until now you have the agent, the editor and files/git/whatever, if you add a fourth panel to that, it becomes pretty cramped at lower resolutions. Fortunately, I do have a 4K monitor, but until now I used to have the editor on one half of it and another window (a browser most of the time) on the other half. So having to use the editor full screen is still slightly annoying.
Of course, that's only the default layout. I'm not familiar enough with Zed, but there's probably a way to change it? In JetBrains IDEs, you can configure panels to sit at the top left/bottom left/left bottom/right bottom/bottom right/top right side and show/hide them with one click (if only one panel on the respective side is shown, it will take up the full space). So you could have files at the top left and the agent panel at the bottom left. And the code editor is of course still the "centerpiece" in the middle.
Luckily it's very easy to change, although a bit unintuitive for new users. You right-click the small icon for each type of panel in the bottom bar, and select where you want it to be docked. Left click toggles the panel into view.
It will also pull more users in. I don't want to be looking at code. I want more of a codex style app where it's easy to shove all my projects in one place and context switch endlessly.
This was also my first impression. But it seems to me the changes are mostly about swapping what panels dock where (left or right) and maybe some additions/tweaks around the AI panels. On macOS these are still the same:
If you opt-in to the new layout, the panels that used to sit in the left dock are now in the right dock. I will give it a try even for classic coding. One can change what panels get docked where from the settings window.Warp launched something similar a week or so ago, but the Zed implementation I find a lot more logical. Will give Zed another try, as I’m overdue for my monthly “maybe I should try this terminal/IDE” itch.
I like Warp but something about it is very opaque and confusing. Maybe it has a learning curve I haven't committed to, or it's just very alpha and evolving often.
Agreed, I exclusively use Warp for server maintenance and ssh'ing into servers, it does that better than Claude itself but the UI is always confusing, especially after their recent changes.
I'm having a hard time adjusting to the Project Panel on the right (and, at least for me, hidden by default) - seems like they're trying to bury the concept of a 'file'?
It's certainly interesting though, and I'll give it some time - the post says "It feels more natural once you've spent a little time with it"
I think it's more that you don't need to look at the actual tree structure all that often, especially when most codebases follow logical and conventionalised structures you remember these structures and can use Command-P to search for files individually to open and the tree structure becomes something you peek at infrequently to ensure that everything is going in it's right place.
You can move it to the other side panel by right clicking the Project icon that toggles the view.
Yesterday, I determined to move to Zed because they weren't pushing this stuff :(
You can disable all AI features in Zed with a single setting: https://zed.dev/blog/disable-ai-features
It doesn’t mean the bulk of their effort isn’t going to AI slopshit.
I disagree. All their "AI slopshit" features are useful, and the fact that you can just.. turn them off and get a great editor regardless is what I like about it.
As compared to the slopshit that you'd rather have them working on to fit your personal needs?
zed's got about 60 billion reasons to add an agent panel
Zed was one of the very-early editors to jump on adding AI. Might want to look elsewhere.
They are at least more tasteful about it, but they do have to keep getting VCs. And no one has ever accused VCs of being smart or technical.
I would love to unleash parallel agents, but I am still checking every single edit while enforcing minimal, stateless, modular code, and I have the AI check in with me before writing the next file.
A lot of times, I find it has incredibly stupid ideas and tends to make the code very messy. I would love to figure out how to stop that from happening automatically.
The upside of checking in on the code, though, is that I can come up with smart directions for the AI from both a product and tech perspective. This is especially helpful when the dumb suggestions add a lot of complexity.
I think it's like when a product person asks for a new feature, or when a founder building their own product selects which feature is smarter to build and how.
I'm expecting we'll likely end back up on agents making PRs, and having to review them. Either that or giving up on quality etc/dealing with very messy code. I've been trying various automated testing/linting/etc strategies, and they only work so well.
That would be a nightmare. One thing is to review a PR generated by a human using AI and caring about the code; another is reviewing wild agents, especially when they make changes everywhere
I'm not excited about it, but the only main way I've been able to discover LLM-isms that sneak in are
1. via seeing them glimpse by in the agents' window as its making edits (e.g. manual oversight), or 2. when running into an unexpected issue down the line.
If LLMs cannot automatically generate high quality code, it seems like it may be difficult to automatically notice when they generate bad code.
> I would love to figure out how to stop that from happening automatically.
AGENTS.md
> AGENTS.md
-- which will be ignored just often enough that you can never quite trust it.
Yup. No matter how much you tell it to keep things simple, modular, crisp, whatever, it generates tons of garbage much too often.
Btw it may be obvious but afaik claude by default only reads CLAUDE.md and not AGENTS.md
And yet still less often than the average developer.
I think the issue is deeper than prompts, agents.md, smart flows, etc. I think the problem is that LLMs are searchers, trained on preferring some results. So, if the dumb solution is there, and the smart solution is not there, they won't spit it out.
I really want to like Zed, but for some reason the way it interfaces with the TypeScript language server is dog slow compared to VS Code and its derivatives.
Check the lsp servers that run when you are editing TS file. I noticed something like 5-6 when editing a TSX file. There was even tailwind thrown in for a good measure (the project has no tailwind). All that buttery smooth 120fps wasted on poor resource management when it comes to LSPs.
Same. Zed is so fast it's shocking. But it was using 2-3x as much RAM as VSCode or Cursor with TypeScript and the language server crashed a lot. Given that work is a TypeScript monorepo, that was a dealbreaker for me.
I'll bet if you point out the issues where this is described and measured, you'll get some eyeballs.
What are parallel agents worth to professional engineers if reviewing the code is a pain (aka non existent) in Zed? Please add proper code review tools (compare with branch|file|revision) and GH pr review tools like IntelliJ!
It has decent git tools, it has had so for a little while now https://zed.dev/docs/git
I'll just leave this here (again): https://github.com/zed-industries/zed/discussions/36604
I remember when Zed's main thing was "collaborative" editing. Not as profitable as AI I suppose.
Zed is probably the best text editor in the last 10 years. It has quirks, but it is insanely powerful and capable out of the box. I don't even bother trying to setup Neovim because of Zed. They let people for PRs for missing vim features for their vim emulation, and its insanely capable.
I hope someday they get the funding they deserve, because it has insane potential. It's why I subscribe to their pay plan, even if I dont use it all the time, I want them to succeed.
I do wish they'd focus on closing the gap to Jetbrains by implementing the QOL features that are missing. I understand they have to do what VC wants to see, but this agentic stuff is so tiring.
I give them a try about twice a year. I write a lot of Rust which should be squarely in their wheelhouse.
This last time I was pleasantly surprised to find they mostly fixed their SSH remote editing support. But then it started truncating rustc inline error messages and I couldn’t figure out how to view the whole thing easily. When you’re just trying to get something done little bits like this can add up quickly. Punted back to Cursor for now.
I don't like the way remote editing works with plugins. IIRC, the remote agent pulls the plugins from the connecting client. I get why it's done like that, but I'd way rather have it go the opposite direction.
I want a setup where I can have an immutable devcontainer with local copies of everything I need to develop 100% offline; dependencies, tools, etc.. Having my local editor pull plugins from a devcontainer for the project seems to make more sense to me.
I didn't dig in too much. Maybe there's a way to make it work somehow.
Agreed. It still feels like Zed is only good for writing scripts.
I give them props for writing a truly responsive editor that's easy to use, and can switch to keybinding sets from other editors. (So you don't have to learn new ones) And I find it works more naturally for multi-file projects than VsCode, and Sublime.
In terms of in-line instantaneous error highlighting, introspection, refactoring, and autocomplete, it's not on the same level as JetBrains.
If Cursor is worth $60bn, how much is Zed worth?
Love Zed for my personal projects at home with Openrouter. But I cant use it at work as I'm forced to use claude code at work and find the VSCode+claude code combo better. I know zed has claudecode integration, but I found it to be very weak compared to zedagent+claudemodel. Has this improved?
Do you mind sharing what your spend is like on Openrouter? I pay for the $200 Claude Code plan just so I never have to think about usage or feel like I'm paying per prompt. But I nearly max out my usage each week and I have no idea what that would cost me on Openrouter. I hear a lot of people are using MiniMax of Openrouter. Is that what you do?
What I want is a stateful file-writing layer that is aware of all clients (aka agents and humans) and their activity. It provides its own locking mechanisms, and prevents agents from overwriting each others work. That way you could have multiple agents operating on the same codebase, without having to futz with worktrees and all that.
tried using zed but when i use codex cli, when agent updates file, it doesn't even register in the "buffer" (why call it buffer? trying to be different?), i have to close the file and open it back up again for it to update the file contents which creates bugs where i edit some other thing in file and save (no "overwrite warning" which i get in vscode) it just overwrites the file....so the editor is just plain buggy.
I loved zed for over 1 year, told for everybody to use it, because it was so fast and great.
But now using claude-code,gemini-cli,codex,etc it just seems less relevant. Just opened nvim with lazyvim and it feels nice, since I'm in terminal anyway it just feels more natural.
Still have zed opened, still like it but I guess honeymoon is over.
I really want to like zed but it claims all over the place to be be built with native windows code but it's the only OS I can't use the OS title bar and this weird menubar that opens on hover?
The claim really falls flat the moment you open it up and it is "native but not as you know it" providing the same UX experience of any other electron junk.
I've used Zed since the very beginning and I remain a fan. Its LLM integration so far has been a lot more pleasant than what I see in others and the editor is perfectly usable without using LLMs.
Its multi buffer and speed sound trivial but using anything else feels wrong now.
I liked the idea of the new layout with the agent thread on the left, it goes hand-in-hand with having multiple threads that are easy to switch between and running concurrently, but I switched back because my file tree disappeared and I couldn't easily see how to add it back
In the new layout, the project panel and git panel are just moved to the right side, so that the agent panel could be on the left, and you could still view both at the same time.
I wonder if you could use something like vLLM and have these subagents max out your local gpu. I’ve been looking for using a local model because I’m tired of rate limits of the cloud and also would really like to make use of my local gpu when I’m working (5090h even if it’s not on my computer I’m typing on.
I just installed Zed last night and enabled vim mode, can't wait to try this!
Funny how Zed's tagline is
At home, I don't use any AI when coding, to keep my brain sharp. But it's clear that Zed's focus is on AI integration because that's where the money's going (seriously, where is the setting to have a different ui icon size vs ui font size). Is there any editor still being being developed and focusing on the experience of coding by hand?> seriously, where is the setting to have a different ui size vs ui font size
Search for font size in preferences.
You'll see a 'font size' under 'buffer' (editor), under 'UI Font', and under 'Agent Panel' to let you control font sizes in all of those places independently.
> Is there any editor still being being developed and focusing on the experience of coding by hand?
Zed lets you hand-edit too! It's fast and decent. vim, neovim, Emacs, Helix, and JetBrains products continue to do that well too. There are still more traditional IDEs/editors than pure AI ones.
You can also toggle AI features off in Zed from preferences if you want to.
ah I forgot a word, I meant the ui icon size. If I bump up the ui font size so that I can distinguish the icons apart on my large monitor, the ui text becomes comically large.
I do use Zed without AI features, it's just a bit of a disappointment (though understandable) since it was originally marketed as just a nice speedy editor.
Have you tried a different icon theme? Some are just easier to see than others. The default icon theme is pretty light.
https://zed.dev/docs/icon-themes
I don't think changing icon size independent of UI font size would be a dealbreaker for many. (I'm quite happy having icons that scale in line with font size, but then I use the Material Icon Theme, which is easy to scan at most sizes.)
It's still a nice speedy editor. It didn't lose any features to make room for AI
Is Zed lacking any feature you need?
A proper git implementation.
I end up doing things in the terminal tab because its faster than the ui or is more clear.
The basics are good but thats about it.
What git features are you missing? We've been adding a ton recently
Just injecting this here: What I've been missing is an equivalent for GitHub's "blame prior revision" feature to quickly follow through the history of individual source lines.
https://github.com/zed-industries/zed/discussions/42583
Thanks for building an awesome product :)
I use zed every day, but the git view is still buggy and unreliable for me. It often chooses the wrong folder---a subfolder rather than the root of my git project---and shows that there are no changes. And sometimes, even when the right folder is selected, it will still incorrectly say there are no changes.
Another bug is that clicking on new files in the git UI doesn't bring up their diff---you need to click them twice to open the file. I'd like to have those files included in the diff view just like other files are.
I'd love for there to be an easier way to get to the "view file history" view---I didn't realize that it existed until I tried searching for it. I'd love a "line history" view like GitLens has, as well.
If you're taking suggestions I'd like to be able to see when I'm over the 72 character limit, last time I checked there was no way to know inside Zed when writing the commit message though I might be wrong. Other than that I think Zed's great and multi-buffer editing is really swell.
Zed is fantastic for coding by hand. The multibuffer editor and 120fps resizing is orgasmic
I'm building Paneditor with that focus. My goal is to close the gap so that (competent) humans can work with code at levels of editing throughput usually reserved for users of LLMs.
I like LLMs, I like Zed, but I turn off the AI features. I rather have Claude or Open Code in a container with only access to a mounted folder, or use a local model.
And Zed lets me do that while remaining fast and minimal.
As for (even more) minimal editors, perhaps just Gnome Edit? Or Kate?
ecode [0] is ridiculously fast -- makes Zed feel like molasses -- and quite customizable. It's still early in development and mostly just made by one guy (who is also developing the GUI framework used), so progress is slow and there are some rough edges, but it has all the important stuff and quite a few niceties too. Really cool project, reminds me of Sublime.
[0] https://github.com/SpartanJ/ecode/
The thing is, they have to monetize somehow. There's a setting to turn all AI features off with one toggle and you're back to an 'editor still being being developed and focusing on the experience of coding by hand'
Doesn't really count, but: https://gram.liten.app/
> ”What cannot be mended must be transcended.”
such a dark and gloomy quote as the mission statement.
Helix?
Vim? Emacs? Sublime Text?
Emacs. ;)
100%. I recently got rid of my lsp-booster and similar kludges because the builtin language server client (eglot) is now fast enough without it, even on large projects.
And if you want AI integration at your choice and control, agent-shell (and chatgpt-shell, which is LLM-agnostic despite the name) are great packages. They’re totally hackable with elisp like you’d expect, which I personally haven’t done a ton with, because I use AI pretty sparingly, but I imagine the crowd here could come up with plenty of ideas for how to program your editor and your agent interface together.
> But it's clear that Zed's focus is on AI integration because that's where the money's going
Do you really think Zed's focus on AI is just about money? You do realize software engineering is in the midst of a tectonic shift?
> You do realize software engineering is in the midst of a tectonic shift?
As an everyday user of AI, both at work and privately, I am not that convinced. The biggest effect I've seen so far is demand for faster work because "everything is faster with agents", but software quality is slowly dropping in software I see around me.
Current AI is very useful as a trivia engine and as a language manipulation tool - i.e. it can quickly extract information from a huge amount of text. But it still sucks when writing new things.
Admittedly, here has been much progress, but it seems to be slowing down. Money is drying out, models are getting nerfed, and only better scaffolding and workflows are making it better. Unless they build 100x more data centers, I don't see models getting significantly better.
> Do you really think Zed's focus on AI is just about money?
Yes? Legitimately curious what other explanation is there here, thats the reason all of these LLM integrations across all software is being pushed.
What will convince people is what they see with their own eyes. Not yet another proclamation that the revolution is happening right now.
Like this.[1]
> AI-assisted coding has become the norm and with tools like Cursor, GitHub Copilot, Claude Code, Codex, we are increasingly letting models touch our code.
... how is it good literaly style to both (1) assume that something is the norm, and (2) use a long intro-sentence to state that something is the norm? Pick a lane—either it is the norm and you don’t need to state it or it isn’t and you need to set the stage. Stating the apparently obvious makes your (their) writing read like a eighth grade paper.
In short I’m agnostic as far as proclamations go. ;)
[1] https://news.ycombinator.com/item?id=47866913
neovim or eMacs are the best text editors as up today.
This is great! I love Zed, but when I came across Superset I stopped using Zed. Maybe no I will go back to it
actually not that great. it's not as intuitive as Superset. Also, I can't manage to get a Claude session in the terminal to become a thread
I've been a Sublime Text user for years, then a VSCode for years. Been trying Zed for the past couple weeks and it has been a good experience.
Becoming more and more useful by teh day. Love to see it.
And yet the file explorer still fails to update when the files actually change https://github.com/zed-industries/zed/issues/30851
Does this work with ACP?
Yes
Zed might have a fighting chance against VS Code now that it's million extensions are quickly becoming irrelevant since agents do all the coding.
What's needed today is a nice way to orchestrate agents and do some small manual edits there and there to the code.
I like Zed but the defaults for tabs (CTRL+Tab and CTRL+Shift+Tab step you through some diabolical tab selector in what is allegedly (but not really) recency order instead of just left/right. It really made me question their sanity.
Love Zed.
Is this any different from a setup where I use a terminal with tabs and splits, running my favorite editor in one or more panes, and several agents (Claude Code and Codex) in several other panes and tabs?
Edit: Although I can integrate an agent in NeoVim, I don’t do it. I want to use my editor solely for that purpose, while the rest (versioning, agentic coding, git client, etc.) is done in the terminal. My NeoVim setup is simple and fast, which is why I prefer it over any other IDE or editor. Especially with the native package manager in the latest version. I also replaced BBEdit by installing Neovide, a GUI version of NeoVim. It starts in a split second and is incredibly smooth and fast. And it’s so enjoyable to work with that I use it as my preferred frontend to Obsidian.
I must be missing something crucial in my obsidian setup. What is Obsidian without its frontend? For me it's just a bunch of markdown files.
I gave zed agent another shot today, and it couldn't run bash commands, so back to opencode I go.
The worktree part is the easy half. Running parallel Claude Code subagents, the bottleneck is never "can they run without stomping each other's files." That's solved the moment each one has its own checkout.
The hard problem is architectural consistency. Agent A renames a type to X. Agent B, in a different worktree, independently renames the same type to Y because neither saw the other's decision. When you merge, neither worktree is "wrong" but the code is incoherent. You need either a shared decision log that every agent reads before starting, or an orchestrator that hands out scope narrow enough that no two agents can collide.
Zed's post is solving filesystem-level parallelism. The harder coordination problem is semantic, and that's where time savings from parallelization go to die.