I wrote a harness leveraging tmux and nvim: https://github.com/wb14123/tcode . It's able to open subagents and even talk with them, all managed by tmux and using nvim as the UI for both the editor and the messages display window.
Once you get it set up, Emacs is a pretty damn good "agent" multiplexer as well. I use agent-shell with Projectile on Doom Emacs as my main workflow these days, and it works very well even if I have 6 projects open or whatever.
Claude and Codex are also both quite good at working with Emacs as well. Depending on your isolation/sandboxing strategy, they can either run commands against your session via emacsclient (a bit scary) or dump elisp in the REPL for you to evaluate. Both are really efficient in terms of giving you a fast feedback loop.
I switched from tmux to Zellij a while back, and lately added a stop hook that sends a terminal ping to the correct tab when the agent is done (and I'm not looking at the tab). It has been pretty convenient so far.
The utility is not obvious, but I've been using herdr for a few weeks after a friend recommended it, and it's been a great tool.
I had about a dozen different terminal windows, each one with multiple tabs, which was becoming a mess to manage; multiple agents and harnesses running that I could only inspect by remote access/VNC away from home.
With this, I can keep things more organized (project workspaces, tabs), because it's backed by a persistent process, I can ssh into the machine, with tailscale, run herdr and see all active sessions, do some debugging and one-off prompts.
Even better, I can do that from my phone and iPad using Prompt/Termius. It gives me the best part of 'xxclaw' harnesses with none of the complexity.
Will people really need a bunch of different agents?
Even using two codex session for different project make me feel a bit overwhelmed, or may be i just old
If you have more than one issue to work on, then you can open an agent for each one and go though them round-robin.
Maybe very early on a greenfield project you only have one known issue (the next feature you're building), but other than that seems like short-changing yourself.
Maybe not when answers will be almost instant, but the biggest gap to me now is latency, not capability. Having to wait for 5 minutes and more for an answer, I can switch to another project / feature / etc. and run work streams in parallel.
I found that a good tool helps a lot once I switched to Github Copilot app. It solved the friction and mental tax for me. I easily manage 4 sessions in parallel on same or different projects while 2 was max in the past. The bottleneck now is only in review and decision making.
Same. I haven't been able to see how people let agent loops run without significant steering and produce good quality software. VS Code with one or two integrated terminals running is fine for me. Or a couple of VS Code instances if I'm working one a couple of projects concurrently. The advantage of VS Code is the code / diff visibility if you like to be hands on.
yeah I also don't really see a use case for me, like do ppl really run that many agents in parallel that they cannot comfortably multiplex them using just a terminal emulator
One interesting thing that I couldn't find in these comments or in their docs (even in the https://herdr.dev/docs/preview/concepts/#mouse-ui section), is that Herdr supports native-feeling scroll. You can scroll with your mouse wheel. That to me is a killer feature over tmux
I've been using this and it's been really nice for a specific use case: running agents locally alongside agents in a remote sandbox environment. Most apps managing this split seem to want to manage your sandbox for you, but we've already built out some nice tooling and infrastructure for that that meets our needs. I don't want to add a new vendor in the mix for our infra that may not exist in a year and that will be harder to integrate with. Combined with some additional tooling I made, herdr lets me easily connect to remote agent sessions, see their status in the sidebar and fluidly switch back to work on my local machine from one interface.
Multiplexing the agents is clearly the first obvious pain point, but the other one I keep encountering after this is visibility: with multiple agents running, it becomes difficult to see what each of them is doing, what program did they call, and where they are getting stuck until they complete or fail.
Is there any information from Herdr about what each agent is up to beyond the output? Or does it just concentrate on orchestration for the time being?
I’ve been using herdr for about a week. There is visibility into which terminals need attention and you can set up workspaces to organize it. I’ve enjoyed it so far though the key bindings are all over the place - likely due to it trying to avoid existing bindings.
I started building my own version of this before I discovered herdr, and while its not quite the same it's improved my development workflow by an order of magnitude. https://jmux.build
Getting a lot of flack in all these comments, I really like this tool. Have been super easy to use and scale to multiple agents. I've had a ton of issues with tmux and copy and paste, this just works. I was using warp terminal, and even working on my own fork of it with it's recent open source status, but this has won me over.
I'd suggest adding a few screenshots on the Github README, otherwise I wouldn't have enough attention to imagine how it looks. I had to go to the website to appreciate your work, otherwise I'm like "what is it?"
has anybody figured out how to get this to open a fixed layout? I'm using tmuxp to do this easily with tmux but would switch to this pretty quickly if it was easy to get my layout over. It's really nice. It's like zellij without the bloat (but with different ai bloat that doesn't seem to drag down performance at all and isn't in your face unless you want it in your face)
How do people use terminal multiplexers together with vim?
Ctrl+B is so hardwired in my fingers for scrolling back one screen that there's no way I'm remapping that one in vim itself. So then you have to remap that in your terminal multiplexer, while at the same time there's a bunch of people saying never change the leader key...
I'm using tmux + zoxide+ https://github.com/joshmedeski/sesh. Then on Ghostty, I have keybinds for my workflow. Cmd+k to open/switch workspace. Each workspace is just a new tmux session.
I use tmux with vim and configure it to use Ctrl-a. Not for vim, but because I started with GNU screen that used this key. For the cases when I need actual Ctrl-a, tmux is configured to send it when I do “Ctrl-a a”.
As a vim user, I just remap C+B to C+A. It's much easier on the fingers too. Issue arises when I ssh somewhere that doesn't have the leader remapped but that's usually pretty rare when I have to vim in a tmux session on a remote host so not really an issue
You might like Circus Chief. It’s a web based UI focused on mobile - local first. Supports Claude Code, Codex, and Gemini. Supports auto retries when you run out of tokens, scheduled tasks etc. It’s been great for me as I juggle a Codex, Claude Code, and z.ai subscription.
does it support a setup where each agent can be in a different SSH session? or must they all run in the same place.
it seems to support running a remote herdr over SSH but unclear if it can add remote agents (each agent has its own sandbox where its installed and you first SSH into it and then start the agent there)
I usually have one local clause orchestrating multiple remote Claude in different tmux . And then another orchester and remote vm worker sin tmux for another repo etc...
It gets a bit hard to keep the overview but I don't want to give up my parallelism, your too might help
I was just looking at conductor and was not very jazzed about the fact that it was running the agents directly on the host. Being able to launch from a terminal means this one can (hopefully) run from inside the sandboxing setup I use for coding agents.
It's not so straight-forward. Either you must add hooks the the agent to notify tmux directly or you must use an external tool that polls tmux to determine one of its panes has gone silent and then based on that, send notification to tmux.
The poling requires tmux (not screen nor dtach, as far as I could find). And, silence for N seconds is just that, the poll doesn't know if that really means waiting for input or something else. With agents (like claude) that have a throbber/spinner going while "thinking", silence is a good indicator.
Kitty terminal can be polled for current text and then see if that has changed in N seconds. This would allow not having to depend on tmux which may be preferable to some kitty users. Generally, using tmux always surfaces some annoying problems for me.
Does this mean adding instructions to AGENTS.md saying to end everything with the bell character? Or do harnesses have this in their settings somewhere?
I think generally harnesses have this. In claude code it's `"preferredNotifChannel": "terminal_bell"` in the settings.json, pi and opencode looks like you have to either add a hook yourself somewhere or use an extension.
Depends on the harness I imagine. If there's some sort of "post run" hook I'm sure it can be added there. Or, if the harness is open source, a PR to add it would work too.
I wrote a harness leveraging tmux and nvim: https://github.com/wb14123/tcode . It's able to open subagents and even talk with them, all managed by tmux and using nvim as the UI for both the editor and the messages display window.
Once you get it set up, Emacs is a pretty damn good "agent" multiplexer as well. I use agent-shell with Projectile on Doom Emacs as my main workflow these days, and it works very well even if I have 6 projects open or whatever.
Claude and Codex are also both quite good at working with Emacs as well. Depending on your isolation/sandboxing strategy, they can either run commands against your session via emacsclient (a bit scary) or dump elisp in the REPL for you to evaluate. Both are really efficient in terms of giving you a fast feedback loop.
Please share any useful guides and/or resources for agentic development with Emacs!
If you like scary, there are also emacs based agents like gptel which support tools like "eval-elisp".
A kitchen sink is almost indistinguishable from a sandbox, right?
... right?
I switched from tmux to Zellij a while back, and lately added a stop hook that sends a terminal ping to the correct tab when the agent is done (and I'm not looking at the tab). It has been pretty convenient so far.
Added an example here:
https://johnny.chadda.se/zellij-stop-hook/
The utility is not obvious, but I've been using herdr for a few weeks after a friend recommended it, and it's been a great tool.
I had about a dozen different terminal windows, each one with multiple tabs, which was becoming a mess to manage; multiple agents and harnesses running that I could only inspect by remote access/VNC away from home.
With this, I can keep things more organized (project workspaces, tabs), because it's backed by a persistent process, I can ssh into the machine, with tailscale, run herdr and see all active sessions, do some debugging and one-off prompts.
Even better, I can do that from my phone and iPad using Prompt/Termius. It gives me the best part of 'xxclaw' harnesses with none of the complexity.
Couldn't you just use tmux for the remote access?
Will people really need a bunch of different agents? Even using two codex session for different project make me feel a bit overwhelmed, or may be i just old
If you have more than one issue to work on, then you can open an agent for each one and go though them round-robin.
Maybe very early on a greenfield project you only have one known issue (the next feature you're building), but other than that seems like short-changing yourself.
Maybe not when answers will be almost instant, but the biggest gap to me now is latency, not capability. Having to wait for 5 minutes and more for an answer, I can switch to another project / feature / etc. and run work streams in parallel.
Dang, that's not too much context switching for your brain?
I found that a good tool helps a lot once I switched to Github Copilot app. It solved the friction and mental tax for me. I easily manage 4 sessions in parallel on same or different projects while 2 was max in the past. The bottleneck now is only in review and decision making.
Same. I haven't been able to see how people let agent loops run without significant steering and produce good quality software. VS Code with one or two integrated terminals running is fine for me. Or a couple of VS Code instances if I'm working one a couple of projects concurrently. The advantage of VS Code is the code / diff visibility if you like to be hands on.
I often have many active. Bug investigations, writing code, reviewing, checking logs after deploys and so on.
I usually have 5-8 at a time
yeah I also don't really see a use case for me, like do ppl really run that many agents in parallel that they cannot comfortably multiplex them using just a terminal emulator
I'm a power user and it's not an issue for me. I just use Kitty with multiple windows and panes. I can jump between using hotkeys with no problem.
One interesting thing that I couldn't find in these comments or in their docs (even in the https://herdr.dev/docs/preview/concepts/#mouse-ui section), is that Herdr supports native-feeling scroll. You can scroll with your mouse wheel. That to me is a killer feature over tmux
I've been using this and it's been really nice for a specific use case: running agents locally alongside agents in a remote sandbox environment. Most apps managing this split seem to want to manage your sandbox for you, but we've already built out some nice tooling and infrastructure for that that meets our needs. I don't want to add a new vendor in the mix for our infra that may not exist in a year and that will be harder to integrate with. Combined with some additional tooling I made, herdr lets me easily connect to remote agent sessions, see their status in the sidebar and fluidly switch back to work on my local machine from one interface.
I finally switched to Zed + Terminal Thread + Tmux.
The terminal agent threads allow me to run in multiple projects. They always run in a project scoped tmux session so they are always on.
And of course Zed does remote via ssh so the agents can run without me needing to keep my laptop on.
Herdr is great, I've been using it for over a month now. Their API is awesome too - you can build tools around it.
Unlike Cmux, Herder will not call out to multiple web services upon start-up.
Multiplexing the agents is clearly the first obvious pain point, but the other one I keep encountering after this is visibility: with multiple agents running, it becomes difficult to see what each of them is doing, what program did they call, and where they are getting stuck until they complete or fail.
Is there any information from Herdr about what each agent is up to beyond the output? Or does it just concentrate on orchestration for the time being?
I’ve been using herdr for about a week. There is visibility into which terminals need attention and you can set up workspaces to organize it. I’ve enjoyed it so far though the key bindings are all over the place - likely due to it trying to avoid existing bindings.
Nothing beats tmux + iterm 2 integration. Its really my terminal.
Can you say more about what it gives you over tmux alone? I've assumed it's just native tmux tabs
I started building my own version of this before I discovered herdr, and while its not quite the same it's improved my development workflow by an order of magnitude. https://jmux.build
I am going to give them a try side by side!
How does it compare with https://github.com/agent-of-empires/agent-of-empires
Getting a lot of flack in all these comments, I really like this tool. Have been super easy to use and scale to multiple agents. I've had a ton of issues with tmux and copy and paste, this just works. I was using warp terminal, and even working on my own fork of it with it's recent open source status, but this has won me over.
if you prefer to self host but with a gui there is https://github.com/kdlbs/kandev
nice, not dissimilar to what i built for https://github.com/storozhenko98/beehive
I'd suggest adding a few screenshots on the Github README, otherwise I wouldn't have enough attention to imagine how it looks. I had to go to the website to appreciate your work, otherwise I'm like "what is it?"
has anybody figured out how to get this to open a fixed layout? I'm using tmuxp to do this easily with tmux but would switch to this pretty quickly if it was easy to get my layout over. It's really nice. It's like zellij without the bloat (but with different ai bloat that doesn't seem to drag down performance at all and isn't in your face unless you want it in your face)
How do people use terminal multiplexers together with vim?
Ctrl+B is so hardwired in my fingers for scrolling back one screen that there's no way I'm remapping that one in vim itself. So then you have to remap that in your terminal multiplexer, while at the same time there's a bunch of people saying never change the leader key...
Curious what vim users especially do about this?
I'm using tmux + zoxide+ https://github.com/joshmedeski/sesh. Then on Ghostty, I have keybinds for my workflow. Cmd+k to open/switch workspace. Each workspace is just a new tmux session.
>Ctrl+B is so hardwired in my fingers for scrolling back one screen that there's no way I'm remapping that one in vim itself.
Give it a month and whatever you remapped will be "hardwired" too.
In any case, no reason to keep Ctrl-B in tmux either, you can remap that just as well.
Change the tmux leader to ctrl-space, problem fixed!
I'd recommend using Page Up and Page Down, but this seems like a much better solution.
I use tmux with vim and configure it to use Ctrl-a. Not for vim, but because I started with GNU screen that used this key. For the cases when I need actual Ctrl-a, tmux is configured to send it when I do “Ctrl-a a”.
As a vim user, I just remap C+B to C+A. It's much easier on the fingers too. Issue arises when I ssh somewhere that doesn't have the leader remapped but that's usually pretty rare when I have to vim in a tmux session on a remote host so not really an issue
This is way to complex... Why don't just use some harness which manages all that and give u a good UI?
I have been actively searching for such a thing for months without finding it.
Some are OS X only.
Some will not work well with VMs/isolated agents.
Currently using emdash, which is going in a great direction but somewhat new and buggy still.
Nimbalyst does a pretty good job. Some gaps but I can manage multiple agents across multiple projects well enough.
What is your top chart? I am using GUI harnesses from model providers and not happy with any of them.
You might like Circus Chief. It’s a web based UI focused on mobile - local first. Supports Claude Code, Codex, and Gemini. Supports auto retries when you run out of tokens, scheduled tasks etc. It’s been great for me as I juggle a Codex, Claude Code, and z.ai subscription.
https://github.com/ferrislucas/Circus-Chief
Full disclosure: it’s my project
for example…?
I don’t believe any current harness offers this level of organization and running bare shell sessions inside.
OpenCode web?
this works so seamlessesly inside emacs out the box without a dedicated app
https://github.com/elij/macher-agent
does it support a setup where each agent can be in a different SSH session? or must they all run in the same place.
it seems to support running a remote herdr over SSH but unclear if it can add remote agents (each agent has its own sandbox where its installed and you first SSH into it and then start the agent there)
Same question.
I usually have one local clause orchestrating multiple remote Claude in different tmux . And then another orchester and remote vm worker sin tmux for another repo etc...
It gets a bit hard to keep the overview but I don't want to give up my parallelism, your too might help
Interesting, how is this orchestration set up?
One way is to use something like https://shellbox.dev - you start new boxes and manage it all through ssh, icl agent running inside
Each work pane gives you a new terminal session, you can ssh into a remote and it will be kept alive. Should work fine.
afaik, you should be able to use named session to achieve this: https://herdr.dev/docs/persistence-remote/
I'm interested but it doesn't look like it'd be very mobile-friendly.
I need a shepherd for my terminals.
Nice, another ghostty spawn. One of these days I'll install one :3
Tried this for a couple days, but conductor.build is way better IME. Running _in_ the terminal is a flex, but doesn't actually bring any advantage.
I was just looking at conductor and was not very jazzed about the fact that it was running the agents directly on the host. Being able to launch from a terminal means this one can (hopefully) run from inside the sandboxing setup I use for coding agents.
Running in the terminal means I can access it via ssh on my phone (& tailscale). Do any of the other solutions let you do this?
I'm using `opencode --web`. Running over HTTP means I can use it comfortably on my phone with Tailscale.
Check out paseo and similar tools to that. It has a lot of features but you'll find its mobile client is much better than opencode's web view.
conductor.build is Mac only ..
mac only. closed source.
How do you use multiple agents in one project? Isn't there race conditions? or is this for 2 separate projects at once?
Git worktrees
Oh. You're a genius. It's the same as 2 people working on one project. Maybe I'm dumb.
No genius. Just too much time on this site :)
Just use tmux no?
This works much better than tmux.
How about running this inside tmux?
How would you know if a tab that isn't frontmost is waiting for input or complete or whatever?
It's not so straight-forward. Either you must add hooks the the agent to notify tmux directly or you must use an external tool that polls tmux to determine one of its panes has gone silent and then based on that, send notification to tmux.
The poling requires tmux (not screen nor dtach, as far as I could find). And, silence for N seconds is just that, the poll doesn't know if that really means waiting for input or something else. With agents (like claude) that have a throbber/spinner going while "thinking", silence is a good indicator.
Kitty terminal can be polled for current text and then see if that has changed in N seconds. This would allow not having to depend on tmux which may be preferable to some kitty users. Generally, using tmux always surfaces some annoying problems for me.
Someone else suggested the bell character. But you can also just set tmux pane names or color. Which you can also do from your agent harnass.
That's only half an answer.
You can achieve this with features that Tmux already have. Glued a tool that works:
https://gist.github.com/Enoch2090/5026c417f86ff6ff4fbe30c22b...
You use an agentic harness which terminates with bell character[0]. This lights up tmux, clearly indicating that it's done.
[0]: https://en.wikipedia.org/wiki/Bell_character
Does this mean adding instructions to AGENTS.md saying to end everything with the bell character? Or do harnesses have this in their settings somewhere?
I think generally harnesses have this. In claude code it's `"preferredNotifChannel": "terminal_bell"` in the settings.json, pi and opencode looks like you have to either add a hook yourself somewhere or use an extension.
Depends on the harness I imagine. If there's some sort of "post run" hook I'm sure it can be added there. Or, if the harness is open source, a PR to add it would work too.
I added sounds: https://github.com/gpurkins/waiting-for-claudot which has become a bit pavlovian. This along with notifications in iTerm is good enough to juggle some claude sessions.
At least one of us needs to up our dosage.