Should Zed's AI Assistant (optionally) become an Agent?

[stippi]

Issue #17681 was closed as unactionable with the suggestion to open a discussion. I'm hoping to gather opinions from core Zed contributors on this topic.

There are different ways of working with AI on code. The approach that Zed is currently taking doesn't trust the AI with a lot of responsibility: You manually give the AI context (slash commands help accelerate that), it can suggest changes which you have to approve. On the other extreme you have autonomous agents such as Devin that don't give you a lot of visibility into what the AI is doing or ways to influence it.

A sort of middle ground, that I personally find very effective, are tools such as Cline. They keep the high-level chat interface, but provide the AI with tools they can use autonomously within an agent loop in-between user messages. The user is given full transparency into what the AI is doing and "thinking". Each tool can be set to auto or manual approval. The most important tools are list-files, read-files, update-files, create-files, execute-command, and search. The key difference to the Zed Assistant is that the AI can create its own context and work on a (sub-)task autonomously. Whenever it is done or needs input, the user can add another message to the "chat", thereby allowing to iterate on the work.

It would be great if Zed offered this workflow at least as an option within the AI Assistant panel.

[aevsai]

@stippi That would be actually great! I think that it would be nice if zed would provide api for listed actions and users would be able to define some llm workflow via config like tasks maybe? Then we would be able to make both: ai agent with llm model in loop and ai workflows where all steps are predefined. It looks like reasonable approach that would give users ability to make different agents and workflows as we want and maybe share it like extensions?