Agentic mode for AI Assistant

[aevsai]

Recently i tried Cursor agentic mode and now want that in Zed (btw I am not the only one with such request). I found some other topics that may be related to this discussion, but I think it would be right to make and discuss feature proposition in correct category.
Some of related issues:

• Should Zed's AI Assistant (optionally) become an Agent? #23396
• AI Assistant should have embedding context of the entire codebase #15924
• ZedAI Multi file edits like Cursor's composer #16562
• Allow to hide the thought process of "thinking" models in the assistant panel #22373
• Apply AI-suggested changes directly from panel #22903

All discussions comes to this list:

• Make assistant better understand codebase context, e.g. gather context automatically depending on request like cursor does. In order to do this @stippi offered to implement tools like list-files, read-files, update-files, create-files, execute-command, and search for tool call, so agents could gather contexts themselves
• Document "Suggest edits" and /workflow, so users would understand how to enable it and what to use (e.g. /workflow is replaced with "Suggest edits", but there is nothing in documentation about it or about when enable_experimental_live_diffs for assistant would become part of stable build and would it be enabled by default?
• Add agentic mode and ability to make custom workflows that would be able to use functions above.
• Hide chain of thoughts with spoiler (config depends on model)

I started to gather requirements to make a plan for realisation and found out that there are questions that should be answered at first:

• How do we spawn new chat with agent? Some other editors have separate "agent" mode, but I am not sure that it is the best way. I actually think it is better to be able to attach agent for existing chat/context like /agent {name} or add selector near the send button or near model selector?
  
• Should we have some way to describe custom agents and workflows? Maybe some json-based config like tasks?
• Should we explicitly define set of functions for tool call?
• How should we approach privacy? Should we limit what agent can do: should we hide env files? should we limit agent to project folder? for something local like llama it is ok to read secrets, but for cloud llms there definitely should be some restrictions
• What key bindings should be available?

[LivioGama]

https://liviogama.notion.site/Building-Cursor-Product-Manager-3-0-18dfec27aed880f59ae8cd8318cd50bf?pvs=4

[aevsai]

@LivioGama thanks for ideas, though it would be more valuable if you summarise provided content and highlight main idea as it hard to keep discussion based on screenshots from chatgpt

[LivioGama]

The whole point is to have the reasoning and the details. The summary is already the first message.

[notpeter]

Do not tag Zed team members into new discussions. Definitely don't tag more team member if they (checks notes) fail to respond within 1 business day. This is notification spam and a violation of the Zed Code of Conduct. Consider yourself warned, don't do this.

Separate of the notification spam, you are not entitled to an immediate response as to the future direction of Zed and whether something is our roadmap. Obviously the topics raised are features we we are considering, but we don't pre-announce features before they have been designed/implemented. If you would like a glimpse at what is coming in the near term (weeks) checkout the #ai Zed channel notes, which is publicly accessible.

[itsaphel]

[aevsai]

I copied discussion into correct category with feature proposition and commented that we should discuss it here #23396 (reply in thread)

[kdcokenny]

Hey, so there's a company competing with cursor in the VSCode realm and pretty much all they do is find the best open source ai coding software and integrate it seamlessly into the IDE. It's called PairAI. Do you think an approach like this may be the best initial approach so that the Zed team doesn't need to build these tools from the ground up?

[aevsai]

Hi, that sounds great, however I am not sure about Zed's policy on bringing such dependencies into project and moreover I can't find PairAI repository or website to check what this project is about. Can you please attach a link to mentioned project?

[Cheirnomus]

The idea of selecting agent modes sounds reasonable. The mode I'd like to see is one where I'm able to provide a prompt template for a complete app structure and then set a mode that then places the results generated into project directories and files automatically.

I have tried to convey that concept and mode in #21259

So far, I check back with Zed but have not seen any progress in the AI department to speak of. I tested my template with WindSurf and it was able generate the project structure. Still I'd like to see Zed lead the charge in the AI IDE arena.

[aevsai]

Thanks for great example of why different agents are good idea, maybe you can provide a repo with example of your usecase? It would be easier to test such thing in zed if we would sync expectations

[Cheirnomus]

My usecase is straightforward. I spend the bulk of my time honing my prompts and get back great results in a great text editor. From there, I just want to run the code ASAP and not have to copy and paste it section by section from the Assistant Panel and also have to build project directories and files that I already asked for in my initial prompt. I just what to be productive on the full implementation loop and work fast not slow.

[savoygrizzly]

After trying out cursor in agent mode I can 100% say this makes the whole things much faster and efficient. I also don't have to spend 5minutes figuring out which small blocks of changes goes where. Zed needs an AI overhaul asap, I have used it for a year but unfortunately I will now use cursor for projects on which I work alone. Zed collab is great, Zed Ai... not so much.

[aevsai]

Small update with a list of useful resources:

• Zed Context Servers – Zed supports context servers based on the Model Context Protocol (MCP), allowing data and tools to be retrieved within a request. Currently, Zed supports only prompts.
• Building Effective Agents – An article from Anthropic about building AI agents and workflows.

At this point, it seems reasonable to extend MCP support for tools and add an MCP tool server as part of Zed. This also aligns with tasks in the #ai channel regarding context inference.

[stippi]

Hi. Tool support in the MCP client would be great. However, a direct integration would allow for a much smoother UX, like seeing loaded files, and suggested edits that can be approved, tool use approval, all directly in the Assistent panel. Also UI for rolling back a batch of changes and so on.

[aevsai]

Can you give an example of how direct integration would be different from MCP? If you want to have control over files that agent would use like be able to confirm files or something is sounds to be fine with MCP:

1. agent request list of files
2. choose what it wants to include into context
3. you review and confirm ->
4. the rest of agent flow

I mean MCP server would still be a part of Zed and probably just would be connected to Zeds API's like "Find in project". Or I miss something about the way MCP works?

[stippi]

As far as I am aware, the MCP server has no control over the UI of the MCP client. So while the MCP client can be aware of what tools the LLM calls (name, parameters), it cannot know what those tools do and how best to visualize them. Or even visualize a cumulative effect of the tools. As an example, the agentic assistant might want to explore the code base. It may read a bunch of files, later some more files, all of which it now has in working memory. A direct integration can have a dedicated UI area showing loaded files.
Another consideration: An MCP client is bound to the chat template. It has to directly forward tool calls and tool results as chat messages to the LLM provider API. This forces the agent loop to accumulate chat messages. It cannot rewrite history when it doesn't know what the tools (in the MCP server) do. This could mean multiple versions of a file are in the context window. An agent loop that is not forced to use the chat template can be more efficient.

[savoygrizzly]

I second this.
Zed is awesome in the way it is designed, but is falling behind in terms of AI. The tab mode is much needed (should come out shortly from what I gather), as well as an agent mode.

[kerma]

Here's an example neovim plugin which gets the workflow I'd be interested in having in zed: https://github.com/dlants/magenta.nvim

It seems the basics are already in place in zed. I'd start by giving the agent LSP as a tool, and allow users to configure additional tools as needed. Another great example of how this has been solved is in goose extensions - https://block.github.io/goose/docs/getting-started/using-extensions
... which itself could maybe be leveraged in order to not re-invent the wheel? Or at least it maybe helps in fleshing out the POC.

[shawnyeager]

Big upvote for goose integration.

[aevsai]

Quick update: from what i saw in recent code: team working on bringing tool calls to assistant, but still, no official announces or updates

I'd say that main goal for us as community rn is to provide clear requirements/expectation on agents mode features and how it should work alongside with usual assistant, bc agents probably would be based on suggest edits feature that right now works as follow:

1. take file and user message
2. insert them into template
3. feed into llm
4. parse result and display with live diff

So to make custom agents there is no need in some additional selectors as i've suggested before, i would say that it could be achieved with slash command to include agent promp. However prompt template for live diff should not oblige you to use it only with attached tab or file.

[aevsai]

Found ui prototype for assistant2 https://zed-prototyping.vercel.app/, looks very cursor-like in #ai notes called Chat with codebase, there is a bunch of tasks in progress

[aevsai]

Roadmap shows that its in early progress https://zed.dev/roadmap

[KorigamiK]

perhaps it's already mentioned somewhere else, but can you share who's responsible for creating and updating https://zed-prototyping.vercel.app/ ? I love the concept and was wondering if it is opensource?

[aevsai]

I suppose its the work of https://github.com/danilo-leal, he is designer at zed

[Aeolun]

Notes on AI channel indicate a target date somewhere early April, which seems a bit ambitious, but would obviously be very nice if hit.

[Aeolun]

Agentic mode would be (even) nices, but for now I’d be plenty happy with a context server that finds/adds files relevant to the current chat.

I’m going to see what I can do with MCP and slash commands.

[aevsai]

I tried to make an MCP extensions for project structure (aka ls -laR from terminal), but couldn't find the way to get current project path. I'm not sure is extensions api even support this rn. Share details pls, if you would find the way to do that.

[Aeolun]

I was planning to specify the path manually in extension settings for now, but I suppose feeding the context server the project path is not an unreasonable PR to submit either.

[btebbutt]

Quick update: from what i saw in recent code: team working on bringing tool calls to assistant, but still, no official announces or updates

I'd say that main goal for us as community rn is to provide clear requirements/expectation on agents mode features and how it should work alongside with usual assistant, bc agents probably would be based on suggest edits feature that right now works as follow:

1. take file and user message
2. insert them into template
3. feed into llm
4. parse result and display with live diff

So to make custom agents there is no need in some additional selectors as i've suggested before, i would say that it could be achieved with slash command to include agent promp. However prompt template for live diff should not oblige you to use it only with attached tab or file.

Not sure if there has been an update since this comment . But I see a lot of comments talking about auto-context as done by cursor - I actually hate this feature from cursor and is the reason I have recently moved to zed. Cursor doesn't show what context is actually being used and therefore you lose complete control as to what is fed to the llm. They don't show the token count and therefore I believe its cursors way to save money given they are charged by the token but users are charged by the message.

Anyway , this auto context should be easily possible through the current zed model context protocol. The agentic mode within already exists through the in line prompt as it grabs the context from the right panel and applies it. The only change that would be great is being able to segregate the model applying the code vs the context model as right now it is locked to be the same in both instances.

The example postgres mcp is great , but is it currently possible for the tab (mcp) command to be something that the the model needs to interact with ? E.g. for the postgres example , what if the query is determined by the llm rather than at run time.

Thanks!

[savoygrizzly]

Cursor's context doesn't matter this much since in agentic mode it can read the tree and read files. So it essentially follows the thread if you instruct it properly.

[btebbutt]

Cursor's context doesn't matter this much since in agentic mode it can read the tree and read files. So it essentially follows the thread if you instruct it properly.

It does matter, because yes as you said it can read directory trees , and read files. But there is a reason cursor don't show token counts. They compress everything, probably using a reranker or small but fine tuned llm. Their goal is to reduce the token count as much as possible since a request of 128k tokens is the same as one with 1 token for their pricing.

This usually is fine, but it's often not and is the reason why it's so bad with with large code bases where code linkage is not straight forward. They should have an option to explicitly choose to have the agent see the full context rather than restricting it to this and hiding the token counts.

[waleed1892]

@maxdeviant Is there any expected release date for assistant2?

[probably-neb]

Please do not ping Zed maintainers asking for information that is either already publicly available, or is not publicly available for a reason. When there is something we feel is ready to be shared publicly, we do.

[tlvenn]

In case this was not on your radar already, I believe CodeCompanion for Neovim is a very nice approach to agentic workflow for a code editor. Probably a thing or 2 to take inspiration from.

[t18n]

3 years of Neovim and I am moving my entire workflow to Zed. All work amazing, especially the keyboard-base workflow. Agentic mode is the only thing that prevent me from switching to it at the moment.

[jvsteiner]

I am coming to Zed from Cursor. I love the fact that you can really control what goes into the context. I dislike the way Cursor's agent mode does tool call after tool call because it makes me feel out of control. I end up having to wait until the agent has its way with my code base and then make the bad-worse decision of whether to reverse everything and reprompt or make an adjusting prompt and hope it will fix the mistakes it just made. I like that Zed kind of forces you to look at the diffs, and consider what you are doing.

With all of that said, I find myself using both editors in the same code base. I use Zed to ask questions, and make plans, and compose detailed requirements. Then I cut and paste the result into cursor as a prompt and let the agent try. The reason is simple. For large edits across multiple buffers, manually adding diffs is just too tedious. Maybe even more tedious than just writing the code in some cases. It will be faster, sure, but is not fun. It feels like the tool is giving the dirty work to me. I could care less if there is an agent mode, that recursively interacts with the code, but being able to apply suggested edits across multiple buffers is a must have. I hope it comes soon!

[stippi]

Hey, I completely understand where you are coming from. But you are doing something wrong with your prompts. Neither should you have to prepare a detailed implementation plan, nor should you helplessly watch the Agent make a mess of your code. Instead, you need to tell the agent to discuss with you first, make a high-level proposal in the chat first, that you can react to and give feedback on. Only once you have a clear understanding of what the AI would change, which means the AI also has a shared understanding of what you want, then you tell it to go ahead and do the changes.
Also, I do like how in Zed, the AI "suggests edits", which means there changes are first "virtual", and you can apply them only once you like them.

[jvsteiner]

I do that, of course - and when I say "prepare a detailed implementation plan" I mean I do that with the AI - in Zed now, where I prefer doing it, but previously in Cursor, where I found the agent to be very eager, and I would have to prompt to hold it back. I guess I could use the non-agent Chat to do that, and then copy paste the plan into the agent window, because otherwise I find myself tiring of writing the instructions, "don't make any changes yet" or "make only the targeted edits I actually asked for for now" . Yes I use .cursorrules, where I have used different approaches with varying degrees of success. I turn off the yolo mode, and react to linter, etc. Anyway the main point is I feel like Zed gives more control, which I like.

[maxdeviant]

We are now accepting signups for the Agentic Editing beta!

[aromeronavia]

I want to join the beta!

[DavidBM]

I also requested access.
This is the feature that would make me pay for any Zed services and finally make me abandon Sublime.

[cccccroge]

holy shoot this is awesome!!!

[savoygrizzly]

This seems really great !

The only things I found didn't work quite perfectly is that I am unsure of whether zed maintains an index of the project. That probably would help with naming conventions and files locations.

Otherwise I really like the review feature the way it's made.

[DavidBM]

Trying it right now and... amazing. This is a game changer!
Looking forward to have it working with open AI gpt 4.1 with its 1M context window.

[savoygrizzly]

So quick feedback on the agent Beta.

It works really well, drag and drop is also in pre, which was the biggest issue up till now.
Is there a shortcut to add selection to the agent planned ? This would gain huge amounts of time. Also one improvement would be to add instructions in the underlying tools descriptions so that the agent tends to split files more often and doesn't go on to create 500+ lines monsters.

My bigger issue now with zed is with its tab completion. I like that it's not as obnoxious as cursor, however it seems like it doesn't "know" the codebase. Having the project indexed if it's not being done already would help. The tab completion also doesn't seem to always follow the coding patterns.

[josser]

When using the local LMStudio with the Qwen model, the agent doesn’t apply code changes—it just proposes them as text.
Is it something expected or I'm doing something wrong?
Sorry if not correct place to ask.

[awoimbee]

I find each model behaves differently, for example Mistral Codestral just returns [{"name": "read_file", "arguments": {"path": "project/lib/module/live/page_live.ex"}}]

[srihari-flex]

when multimodal for agentic editing