Parallel Agent Execution in Zed : Coding at the Speed of thought.

[cs50victor]

Following up on #33132 from June - I've been thinking about the architectural concerns raised and believe there's a lighter-weight path to multi-agent support that could ship sooner.

What's changed since June:

• The agent ecosystem has matured significantly
• More users are hitting the "waiting on AI" bottleneck daily
• Patterns for effective multi-agent workflows have emerged from heavy usage

Minimal viable approach that sidesteps the concerns:

Instead of deep architectural changes, we could start with a simple response queue:

• When you send to Agent A, auto-switch to Agent B if it has a ready response
• No complex delegation UI - just Alt+Send to stay on current agent
• Use existing git worktrees for natural isolation (no new checkpoint system needed)
• Single input bar that can target multiple agents (like addressing multiple people in Slack)

This preserves Zed's "speed of thought" philosophy - you're never waiting, always reviewing or
writing. The implementation touches mainly acp_thread for queuing and adds a thin orchestration
layer.

Why revisit now:

• Users are naturally opening multiple Zed instances to work around the single-agent bottleneck
• Massive productivity gain with parrallel agents.
• Small initial implementation could inform the eventual "proper" architecture

Would the team be open to exploring this lighter approach as a stepping stone? I understand the
previous concerns about doing it right, but even this minimal version would be transformative for
daily workflows.

Timeline-wise, could this fit into early 2025 roadmap? Happy to help prototype or test.

[japboy]

I really want this functionality ASAP. It would be great if the team has some progresses about this feature.

[sbscan]

Perhaps the only reason to keep Cursor.

[antont]

Perhaps the only reason to keep Cursor.

Yep I was now testing Zed again, getting tired of the UI stalls in Cursor. Launched an agent to do one task, then was starting a new one, but couldn't find my first one then anymore.

I figure I'll keep using Cursor as the main tool and will run some single agent for side tasks in Zed occasionally.

[buzzvolt]

does mysti help here?

[buzzvolt]

claude code now has agents swarm (and subagents)

[Kaidstor]

waiting 👍

[DayanLeksonoPutro]

yes, waiting

[tristanbiesecker]

Would love this feature!

[wlamnorman]

Would live this as well!

[alexanderdamasoc-oss]

I will stop using Zed until this feature is complete unfortunately

[franciskafyi]

We're working on this right now. If you'd like to join the Beta, sign up here: https://forms.gle/tXyicuaftR9ETo1y8

I've created an issue as well for good measure: #51426

[cs50victor]

Signed up. Can’t wait to try this. Would love it if Zed supports remote agents, bringing your Claude / Codex subscriptions and customizing remote environments (Sprites.dev might be a good service to either whitelabel or copy from as it has stateful sandbox environments with native fs checkpointing). Excited for all the possibilities.

[cod3licious]

I think it would be great if the "one git worktree per agent" would be optional - sometimes I just want to ask a second agent a question about the codebase or let it make a plan for something; a separate git worktree sounds overkill for this?
Or maybe I'm just not familiar enough with git worktrees, but I just hope that it will be seamless to let two agents work on the same branch / MR without me having to understand the internals of how the agents are interacting with my codebase and git 🫣