All of the handful of tools in this space that I've tried have modeled an overlap/coupling between project, directory, and machine. For my workloads specifically I think these should be organized differently. Something like top-level projects that have child task/goal/sessions with each one fanning out to different agents running potentially in different worktrees, directories, and even on different machines. E.g. I have a lot of notes and other context stored locally. On a different machine I have repositories checked out and build tooling installed. Frontend and backend exist in different directories but can both be relevant to a given unit of delivery. Are there any agent orchestrators that can model things that way?
There's also sometimes options that try to have a tighter integration with a single harness, which honestly seems to be better cause ACP is fairly bare:
Personally, I found Paseo to be very pleasant, except sometimes I had sub-agents kinda hang.
Kepler had a really nice issue tracking integration (GitLab/GitHub Issues, Trello, Linear, Jira) but their UI felt like vibe coded slop, like the core experience doesn't feel very good, but they keep adding shit that nobody needs like a graph visualization of what the agents are doing, instead of letting me drag the tasks around the Kanban board myself or tell the agents how to do that and how to transition the issue statuses. Hopefully it's gonna get better within a few years, still early on in development. Still, the idea of an issue tracker integration is brilliant and all harnesses should consider supporting that, e.g. pick which tickets you want to work on, they automatically create worktrees and agents do the planning/research step, presenting you with plans like /grill-me so you can decide what to work on, alongside having a step for merging all of those worktrees whenever you want when a wave of work is done.
I haven't tried most of them tbh cause my OS install doesn't need to be even more of a dumpster, but I'm surprised at how none of these really seem all that... rock solid?
Like if I open a JetBrains IDE like IntelliJ IDEA / Rider I know that despite the crappy resource usage and whatnot, they have the needs of writing, refactoring, debugging and building code covered. Most of the current harnesses seem to get some baseline of telling agents what to do, but still somehow end up being buggy (e.g. can't peek into OpenCode sub-agents and give THEM instructions last I checked, or pause or alter the prompt, despite Claude Code allowing message exchanges with sub-agents and such), and many others are just clones of what already exists.
At the same time the desktop / web apps shipped with harnesses like OpenCode (if there's one at all) are very clearly insufficient and a step below Claude Code Desktop and Codex (ChatGPT app) in usability (like that move to horizontal tabs like browsers have, come on how am I supposed to navigate when there's like 30 of those? or what about an auto mode switch within the UI), so obviously we do need something as an ADE.
Personally I wonder if half of these projects will die off in like 5 years, or whether they'll get more polish or differentiate themselves more, and which will try to push more for cloud-only stuff, like how RooCode essentially dumped their VSC plugins and went all in on the cloud.
My macOS is a dumpster, so I can tell you that all these products do slightly different things. It’s apples and oranges across the board.
Except instead of just apples and oranges, there are several dozen fruits involved, and it takes a month to fully taste and digest any of them enough to decide whether they’re worth eating. And by then, there might be an improved cultivar of one of the fruits you tried before but decided wasn’t good enough.
Love these threads for sharing experience. Using Orca has been a game changer (which I found and started using because I saw somebody using Conductor)
These sorts of management tools really speed up development by automating all the PR management, such as making new worktrees, etc. and then super speed up the multitasking between different parallel feature development.
Excited to see if this one works better for me, and to try all the new alternatives in these comments.
Me too! (except for a new work tree) If I don't gate on the git commits it's too hard to catch the code and algorithmic errors.
I had to add some pretty severe harness restrictions for the disastrous Claude 5 line to get it to follow directions on that (ended up cancelling my sub over that sort of behavior, and the bad use of human language)
I still believe this space to be unclaimed, and I am yet to see who's going to crack the demand and needs of the niches out there.
I don't think we even have a great understanding of the market, the niches, and where it is going.
It's also evolving very aggressively constantly, sticky workflows get replaced dramatically whenever new UX or tooling brings a shift into how things can be done.
I'm very curious if anyone has good research or idea of this space mapped out somewhere, I am trying to build for it too but it's really foggy right now and the North Star moves often.
How's this different to https://paseo.sh/ or https://www.onorca.dev/ which seems to be more feature complete? I personally use and like Paseo. I get to connect to my LLM agents even on other machines, remotely via Tailscale and the mobile app UX is quite good.
I think every single one of us who use more than one agent locally has built something like this. There is definitely a need, and appreciate it being 100% free, but eventually, something needs to be so good and with so much community best practices in it, that it will beat a DIY one. E.g. I usually use Claude Opus 5.5 as the "driver" and let it spin cursor, codex, other claude subagents at will for different tasks (this way it always know what is running where). Simply asking it to break down the backlog into a DAG, show which one needs my input, which one is blocked, publish an HTML page and keep it up to date (and sinked with the built in TODO list) is usually all it takes.
I would go as far as making this a skill perhaps, so I can get the same results on other projects... but not sure what is the benefit of making it a Mac only binary, it can be a simple HTML / jsx, with maybe hooks or a /loop to keep it up to date.
A solution needs to be really better than what I can ask Claude to build in 20 minutes. It's not the "someone else spent time and thought and tokens so you don't have to" which does have some value, it's knowing that whatever I try, it will be a little different from what I want, and the energy to fork it, tweak it, is too high vs just asking Claude to build it for me on the spot...
tl;dr Definitely cool, definitely needed, but the bar for everything (including free and open source) has gone extremely high. Things that would become beloved de-facto "industry standard" tools a la "oh are you still using X everyone is using Y" just a few years ago, are now shrugged as a vibe coded lunchtime project (the new weekend project)
It's impossible to impress anyone anymore. I feel dead inside. We all got accepted to hogwarts and we became desensitized to having magic powers.
Huh. I just got pitched something like this earlier this week by an agent as well.
I'm not the target market for this, but this one has the same limitation as the one that pitched me: Mac Silicon only. No Linux, no Windows, no Intel Mac either.
Sibling comment mentioned Herdr, and Herdr is on all those platforms and apparently has raised $6 Mil.
The other catch for all of these: do you integrate with my custom harness? What do you offer that I can't vibe code myself?
You are right. You can build your own, or you can pick one you like and tweak it as you please.
I think different products cover different needs. I think there's lots of options nowadays and I'm most of the options shine one way or the other, it just depends on the mental model of the user and their goal.
I started my own tool less than amonth ago [1]. I did my first Show HN a week after, I was really proud, I added lots of features that first week just to accomodate my own needs and I was not even focused on this project but I could already see how it really helped me remove a lot of fatigue on managing multiple projects and be capable of manage them all at my own pace. My tool, in particular, was created with a focus on iterating over projects, not agents, and in the human (you) having full control on what you do. Anyway, the thing is that the very same day I posted the Show HN, there was this other project that was also submitted some hours before [3]. This project had more traction and comments already and was the one that appeared on the first page of HN. I haven't tried it, but I'm sure it would be great as well.
Yesterday I had a call with another dev at my organization, and he was also doing his own thing (unaware of the one I did, and unaware of other options available already) and I could already see the overlap with different tools and I even pointed out other projects for him to gather for ideas. I think is fantastic that so many people are exploring this space.
The reality I'm observing, with the latest state-of-the-art models, is that they are amazingly capable models and that less and less I need to even review what they produce. I trust on them more and more. This has deep implications moving forward.
I think we'll see two trends moving forward:
- Devs will build their own tools their own way (not all, but some will)
- Devs will clone an open-source project, and then, instead of forking and pusing upstream, they will just adjust to their own needs and detour from the original project more often than before
Sure enough, some people will just take the convenience of a ready-made project, and not even bother to iterate over your own thing, but when you are one prompt away from fixing this big issue or this subtle detail that annoys you, you don't need a config param anymore, and you don't need to wait days or weeks for somebody to fix, if you can compile the project in your machine, you can just create your own version of the software that works the way you want. You are just one prompt (or 10 or 50) of creating or adjusting to your own preferences.
Industry-wise I think this feels sometimes scary, but I think it is amazing.
Why would I want the terminal coding agents inside of a GUI app? The idea is terrible.
The first point in the explanations below makes no sense, claiming my agents will "keep working" if I close my notebook's lid or disconnect the network.
Some people like to run their agents on different headless servers. Herdr is cool because you can connect via ssh and use herdr on that machine from anywhere. When you close the client or disconnect, herdr is still running on the host.
> Why would I want the terminal coding agents inside of a GUI app?
Because they get difficult to manage with enough of them and multiple projects. The GUI helps them stay organized, especially with multiple harnesses. Herdr is a minimal GUI, and works like a simple tmux, which is why I like it.
All of the handful of tools in this space that I've tried have modeled an overlap/coupling between project, directory, and machine. For my workloads specifically I think these should be organized differently. Something like top-level projects that have child task/goal/sessions with each one fanning out to different agents running potentially in different worktrees, directories, and even on different machines. E.g. I have a lot of notes and other context stored locally. On a different machine I have repositories checked out and build tooling installed. Frontend and backend exist in different directories but can both be relevant to a given unit of delivery. Are there any agent orchestrators that can model things that way?
I built a website that compares 68 of these tools: https://agentmgmt.dev/
Offrun is added.
Nice, would https://vibepod.dev/ also be a fit for the list?
Also maybe https://omnigent.ai/
openrig also seems to fit
reminds me days of coinmarketcap
We already have similar products, like Goose [0], Paseo [1], and T3 Code [2].
[0]: https://goose.ai
[1]:https://paseo.sh/
[2]: https://t3.codes/
Add these to the list:
[3] https://orchestrator.inc/
[4] https://gitkraken.com/kepler
[5] https://emdash.com/
[6] https://superset.sh/
[7] https://www.conductor.build/
[8] https://nimbalyst.com/
[9] https://parallelcode.app/
[10] https://www.jetbrains.com/air/
There's also sometimes options that try to have a tighter integration with a single harness, which honestly seems to be better cause ACP is fairly bare:
[11] https://openchamber.dev/
Personally, I found Paseo to be very pleasant, except sometimes I had sub-agents kinda hang.
Kepler had a really nice issue tracking integration (GitLab/GitHub Issues, Trello, Linear, Jira) but their UI felt like vibe coded slop, like the core experience doesn't feel very good, but they keep adding shit that nobody needs like a graph visualization of what the agents are doing, instead of letting me drag the tasks around the Kanban board myself or tell the agents how to do that and how to transition the issue statuses. Hopefully it's gonna get better within a few years, still early on in development. Still, the idea of an issue tracker integration is brilliant and all harnesses should consider supporting that, e.g. pick which tickets you want to work on, they automatically create worktrees and agents do the planning/research step, presenting you with plans like /grill-me so you can decide what to work on, alongside having a step for merging all of those worktrees whenever you want when a wave of work is done.
I haven't tried most of them tbh cause my OS install doesn't need to be even more of a dumpster, but I'm surprised at how none of these really seem all that... rock solid?
Like if I open a JetBrains IDE like IntelliJ IDEA / Rider I know that despite the crappy resource usage and whatnot, they have the needs of writing, refactoring, debugging and building code covered. Most of the current harnesses seem to get some baseline of telling agents what to do, but still somehow end up being buggy (e.g. can't peek into OpenCode sub-agents and give THEM instructions last I checked, or pause or alter the prompt, despite Claude Code allowing message exchanges with sub-agents and such), and many others are just clones of what already exists.
At the same time the desktop / web apps shipped with harnesses like OpenCode (if there's one at all) are very clearly insufficient and a step below Claude Code Desktop and Codex (ChatGPT app) in usability (like that move to horizontal tabs like browsers have, come on how am I supposed to navigate when there's like 30 of those? or what about an auto mode switch within the UI), so obviously we do need something as an ADE.
Personally I wonder if half of these projects will die off in like 5 years, or whether they'll get more polish or differentiate themselves more, and which will try to push more for cloud-only stuff, like how RooCode essentially dumped their VSC plugins and went all in on the cloud.
My macOS is a dumpster, so I can tell you that all these products do slightly different things. It’s apples and oranges across the board.
Except instead of just apples and oranges, there are several dozen fruits involved, and it takes a month to fully taste and digest any of them enough to decide whether they’re worth eating. And by then, there might be an improved cultivar of one of the fruits you tried before but decided wasn’t good enough.
And tmux with a good config and scripts.
Love these threads for sharing experience. Using Orca has been a game changer (which I found and started using because I saw somebody using Conductor)
These sorts of management tools really speed up development by automating all the PR management, such as making new worktrees, etc. and then super speed up the multitasking between different parallel feature development.
Excited to see if this one works better for me, and to try all the new alternatives in these comments.
I totally keep my agents away from git writes.
Me too! (except for a new work tree) If I don't gate on the git commits it's too hard to catch the code and algorithmic errors.
I had to add some pretty severe harness restrictions for the disastrous Claude 5 line to get it to follow directions on that (ended up cancelling my sub over that sort of behavior, and the bad use of human language)
I still believe this space to be unclaimed, and I am yet to see who's going to crack the demand and needs of the niches out there.
I don't think we even have a great understanding of the market, the niches, and where it is going.
It's also evolving very aggressively constantly, sticky workflows get replaced dramatically whenever new UX or tooling brings a shift into how things can be done.
I'm very curious if anyone has good research or idea of this space mapped out somewhere, I am trying to build for it too but it's really foggy right now and the North Star moves often.
How's this different to https://paseo.sh/ or https://www.onorca.dev/ which seems to be more feature complete? I personally use and like Paseo. I get to connect to my LLM agents even on other machines, remotely via Tailscale and the mobile app UX is quite good.
I think every single one of us who use more than one agent locally has built something like this. There is definitely a need, and appreciate it being 100% free, but eventually, something needs to be so good and with so much community best practices in it, that it will beat a DIY one. E.g. I usually use Claude Opus 5.5 as the "driver" and let it spin cursor, codex, other claude subagents at will for different tasks (this way it always know what is running where). Simply asking it to break down the backlog into a DAG, show which one needs my input, which one is blocked, publish an HTML page and keep it up to date (and sinked with the built in TODO list) is usually all it takes.
I would go as far as making this a skill perhaps, so I can get the same results on other projects... but not sure what is the benefit of making it a Mac only binary, it can be a simple HTML / jsx, with maybe hooks or a /loop to keep it up to date.
A solution needs to be really better than what I can ask Claude to build in 20 minutes. It's not the "someone else spent time and thought and tokens so you don't have to" which does have some value, it's knowing that whatever I try, it will be a little different from what I want, and the energy to fork it, tweak it, is too high vs just asking Claude to build it for me on the spot...
tl;dr Definitely cool, definitely needed, but the bar for everything (including free and open source) has gone extremely high. Things that would become beloved de-facto "industry standard" tools a la "oh are you still using X everyone is using Y" just a few years ago, are now shrugged as a vibe coded lunchtime project (the new weekend project)
It's impossible to impress anyone anymore. I feel dead inside. We all got accepted to hogwarts and we became desensitized to having magic powers.
Huh. I just got pitched something like this earlier this week by an agent as well.
I'm not the target market for this, but this one has the same limitation as the one that pitched me: Mac Silicon only. No Linux, no Windows, no Intel Mac either.
Sibling comment mentioned Herdr, and Herdr is on all those platforms and apparently has raised $6 Mil.
The other catch for all of these: do you integrate with my custom harness? What do you offer that I can't vibe code myself?
You are right. You can build your own, or you can pick one you like and tweak it as you please.
I think different products cover different needs. I think there's lots of options nowadays and I'm most of the options shine one way or the other, it just depends on the mental model of the user and their goal.
I started my own tool less than amonth ago [1]. I did my first Show HN a week after, I was really proud, I added lots of features that first week just to accomodate my own needs and I was not even focused on this project but I could already see how it really helped me remove a lot of fatigue on managing multiple projects and be capable of manage them all at my own pace. My tool, in particular, was created with a focus on iterating over projects, not agents, and in the human (you) having full control on what you do. Anyway, the thing is that the very same day I posted the Show HN, there was this other project that was also submitted some hours before [3]. This project had more traction and comments already and was the one that appeared on the first page of HN. I haven't tried it, but I'm sure it would be great as well.
Yesterday I had a call with another dev at my organization, and he was also doing his own thing (unaware of the one I did, and unaware of other options available already) and I could already see the overlap with different tools and I even pointed out other projects for him to gather for ideas. I think is fantastic that so many people are exploring this space.
The reality I'm observing, with the latest state-of-the-art models, is that they are amazingly capable models and that less and less I need to even review what they produce. I trust on them more and more. This has deep implications moving forward.
I think we'll see two trends moving forward:
- Devs will build their own tools their own way (not all, but some will)
- Devs will clone an open-source project, and then, instead of forking and pusing upstream, they will just adjust to their own needs and detour from the original project more often than before
Sure enough, some people will just take the convenience of a ready-made project, and not even bother to iterate over your own thing, but when you are one prompt away from fixing this big issue or this subtle detail that annoys you, you don't need a config param anymore, and you don't need to wait days or weeks for somebody to fix, if you can compile the project in your machine, you can just create your own version of the software that works the way you want. You are just one prompt (or 10 or 50) of creating or adjusting to your own preferences.
Industry-wise I think this feels sometimes scary, but I think it is amazing.
[1] https://github.com/pausan/agenttik
[2] My Show HN 2 weeks ago: https://news.ycombinator.com/item?id=49720222
[3] Another, more successful, Show HN same day as mine: https://news.ycombinator.com/item?id=49713894
edit: formatting issues and minor tweaks on wording
Congrats on this ship, it's hard to get anything out there!
However, as many other comments have mentioned -- how do you plan to get users in a crowded space?
Why does "and what every account has left." activate the "written by LLM" part of my brain?
Because every bit of this webpage was written by Claude.
I read somewhere that a common thing in text written by LLMs is, in their sentences, putting inanimate objects in the active position.
It occupies the same space in my brain as the clickbait-type "This is the key" and "what you're missing" LLM-isms
Thing I'd want surfaced is per account rate limit headroom before I dispatch, not after I'm stuck mid task.
The obvious question -- which you should spend time answering -- is what does this give that Herdr doesn't have? What's the differentiator?
I don’t really get Herdr it’s terminal with integrated tmux ?
It looks like crap.
Why would I want the terminal coding agents inside of a GUI app? The idea is terrible.
The first point in the explanations below makes no sense, claiming my agents will "keep working" if I close my notebook's lid or disconnect the network.
This is obviously not the case.
Some people like to run their agents on different headless servers. Herdr is cool because you can connect via ssh and use herdr on that machine from anywhere. When you close the client or disconnect, herdr is still running on the host.
> Why would I want the terminal coding agents inside of a GUI app?
Because they get difficult to manage with enough of them and multiple projects. The GUI helps them stay organized, especially with multiple harnesses. Herdr is a minimal GUI, and works like a simple tmux, which is why I like it.
My mental model (and I don't really use it) is it's agent-aware tmux.