I do like the embrace of The Office as a theme because it so accurately represents the dysfunction of all of the agent swarms Iâve seen: Different personalities pursuing their own little goals that are all competing with each other in subtle ways that eventually leads a funny collapse of the outcome you wanted.
Thereâs no denying that LLMs are getting better by the month, but the current wave of LLM office and personal assistants reminds me of the old trend where people were hiring personal virtual assistants from foreign countries to manage their work and interface with people. It seemed like an obvious time saver but every time I interacted with someoneâs virtual assistant it felt like I was playing a little game of navigating hidden structures and communication barriers to get the message I needed to the person hiding behind it all. Then there were the inevitable scheduling failings, missed meetings, dropped emails, and other things that get blamed on the assistant. Iâm getting deja vu watching it all happen again with LLMs replacing virtual assistants and outsourced teams.
Maybe these concepts work for people who are trying to solo dev and who find it interesting to set up and debug little systems for everything they do, but I really donât like working with anyone who surrounds themselves with one of these multi-agent coordination systems as an external shell. Keep it to the internal work and maybe itâs fun for some people.
In areas of applicability where GenAI to now considered to be good/great, the pattern I've observed is a frighteningly rapid progression from "abjectly terrible", to "middling and mostly more trouble than it's worth" and then "good/great".
Agentic coding and image generation are probably the best examples of this.
There's no guarantee that this trend will be seen with virtual assistants, but if it does, we're on the brink of incredible. Dedicated personal assistants, that actually know and understand one's context and work well together, are unarguably hugely appreciated by those who can afford them.
Hey guys, thanks for putting it here, I am Chaitanya I built Munder Difflin, I am here to answer all your questions(except nylonstrung).
For people who haven't tried it: Munder Difflin is a local multi-agent harness that wraps around your existing claude code and codex subscriptions(we literally support almost all harnesses/coding agents).
Simulations are deterministic, they do not consume tokens, infact most of the users(20K+ in a week) say that it has reduced their token consumption due to a benchmarked memory layer acting as a hive mind called mempalace.
Common use cases apart from coding:
1. Create triggers that runs an live agent with your context(Webhooks, slack, scheduled)
2. Almost any kind of automation for yourself(I make it review PRs, send cold emails with enriched context, manage discord, Send myself analytics about how app is doing on email an end to end AI video production and posting workflow in 1 prompt and then some)
It's interesting but It was hard to tell from a quick read if this was a fun game with LLMs or a productivity tool. You could stand to make that more clear.
i asked him why he's pigeon-holing it. and you, too? don't hackers enjoy a little bit of playful ambiguity in their lives? must it be all serious all the time?
Sir, this is the Internet in 2026. We don't allow playful anything unless the intent is clearly labeled and approved by your local decency csar. Please continue to refrain from double entendre, sarcasm, irony, or any other form of misunderstandable communication. Every thought must be of clear intent to the community so we may apply the correct direction of moderation to facilitate the amplification or suppression of thoughts in line with the already accepted and centrally approved community morality. Thank you for your cooperation.
When you say no code leaves your machine, how does code get evaluated by your LLM of choice? Doesn't Claude see your code when you send a prompt just like it usually does?
Nailed it. I constantly get âcold emailsâ that were clearly written by an AI that read 2-3 pages on my web site, my LinkedIn bio, and something else on my social media. They all follow the same formula: try to break they ice with something funny about somewhere I took a photo recently, then segue into what I do for work, then offer to help me do it better - usually by offering leads or CRM improvements.
Itâs such an obviously repetitive recipe that someone must have published as an example somewhere, and now lots of people (or bots) are trying to follow it, all mining for gold that doesnât exist.
most cold emails end up spam, only a small fraction end up being useful. My cold email strategy respects google's limits are very personal and mostly not for selling.
Why I mentioned this here: Everyone knows about cold emailing, this is one of the ways I describe the personalisation and integration. It does, slack, discord, and practically works with anything that has an api key.
I often get cold emails from people interested in my very specific hobby (see website in profile). These are desired and fun.
Mind, they are not written by LLMs. I would consider that spam, outside of very limited circumstances (maybe a non-english speaker using an LLM for translation).
Cool project! Iâm still trying to figure out if this sort of thing is the future, or something that the big AI labs will subsume into their products on their own soon. Or something that wonât be necessary in a few months when LLMs get smarter.
You won't see big AI companies allowing other providers in their native agents/harnesses. I believe this is a great opportunity for open source projects like us.
How is this different from a Hermes agent? Any major pro? Besides the cute graphics.
And why do you need E2E encrypted comms between your own agents? What is that protecting against? To prevent other agents snooping and maybe getting derailed? Or do you support conversations between agents of different users?
Let's say if you are away from your computer and your collegue asks you for a report that's anyways gonna take you one prompt on your agent.
Your Collegue's Munder Difflin agent can directly talk to your Munder Difflin agent through our network, you'd want your messages end to end encrypted.
To get past the uh⌠man I really lack the words to describe the responses that youâve already received, but I guess to avoid them going forward, I would re-label the point your making as being more culturally sensitive. Itâs not asking people to accommodate your faith, rather your culture. It removes the convicting tone and imposition of your beliefs that others wish to derive from your request and instead communicates a desire for representation that isnât âmandatedâ but appreciated.
It probably wonât prevent people from getting worked up over the request but it sheds a lot of potential hate and ammunition.
Perhaps consider reframing the term from referring to the God in which you believe but as merely a reference to just the top agent in its little Munder Difflin universe?
Seems like both a mental model which more accurately represents the meaning of the "GOD orchestrator" term, and solves the meaning disconnect closest to the source that is troubling you, thus being the most effective and efficient solution for everyone?
Ok, I've been running it for a couple hours and below are my thoughts. Please note that I do find it fascinating even if most of what I'm about to say is complaining about the parts I like less.
- Pipelines, not agents. Roles, not agents. I really don't like the idea of defined agents with their own prompt. I want to define roles and spin up N agents with that role. Furthermore I want pipelines "Plan -> Review Plan -> Approval Gate -> Develop -> Code Review + Fix loop -> QA -> Approval Gate -> Merge -> [Ship]". I don't like the work just bouncing around seemingly randomly
- Settings don't seem to save/persist? Or some of them don't. I couldn't let "Michael" spin up agents on "his" own and then randomly he did it even though the setting was still off. Settings has the normal LLM jank I've seen.
- macOS Notifications are broken, they send for any little reason, and then they don't send when you're actually needed. It's like each agent finishing a round causes a notification.
- Speaking of missing notifications, the _most important_ screen to me is the "Ask Me" tab under "Michael", where they ask questions (more on that later) but there is zero indication that anything is waiting for you. You have to dig into it yourself.
- The "Ask Me" tab is great.... when it works. I've had to unstick agents or answer questions they were waiting on answers for
- Trying to be too cute, it was cute for a minute, now I don't care (and I _love_ The Office). I want a more utilitarian view. I want to see questions, plans, be able to inject new ideas, and a small overview of what each agent is doing. I don't need half the screen taken up with a "game ui".
- Why no clear? I don't understand at all the idea of them keeping context. Maybe I'm missing something and I shouldn't be using persistent agents except for more persistent jobs (like Michael's?).
It's an interesting concept, very "Gas Town", and it make me want to write my own that does more of what I'm looking for but I don't have the time (or tokens) currently to take on another project. My current best approach of herdr+6-10 Claude Code sessions feels like it works better than this and keeps me close enough to the decisions I want to make.
I think you and I are cut from the same cloth. Role-based pipelines and coordination with scoped permissions generated at runtime and the ability to communicate is the end game in my opinion, for now at least. Orchestration finally clicked with me when I tried out Gas Town six months ago and I immediately created a system to offload it to the cloud by using a combination of GCP CloudRun + Firestore (for temp file storage and to fill role of beads). Right when I finished ironing out the kinks Anthropic laid down the hammer on how users can use subscriptions and I abandoned it.
Now I am using custom pipeline/messaging control plane that spawns microvms with scoped creds minted at runtime. I had to make my own control plane specifically because everyone is focused on agents with personalities rather than roles and pods of workers. I like everything about Buzz except for that.
I'm slowly inching towards my own control plane because my skill-based workflows (not Claude Code Workflows) are brittle but still work better than anything else I've tried. Enforcing the pipeline has been mostly successful but I'd greatly prefer something more structured.
I've been dipping my toes into the waters of higher-level orchestration for a while now but keep scurrying back to driving CC mostly directly. Any extra layers I've added have just caused too much waste (time & tokens) or otherwise produced inconsistent results. I want to "live" in the decision making space. I want to be consulted for decisions (with context provided so I can context-switch easier), I want to review plans, I want to review the actual UI (when there is any), and I want to be able to easily inject new ideas/tickets into the machine (and probably so way to adjust issue priority).
Thankyou so much for the feedbacks, I'll pick them up in next few versions. If you try it again in a week or so, you'd find a lot of these issues resolved.
- It seems like letting Michael spin up things is the prefered pathway instead of defining a bunch of agents.
- The Ask Me really needs some love. The "X/N" display doesn't make sense to me, I started on "5/5" and as I answered the questions it went to "4/5" -> .... "1/5", I'm not sure if the questions were asked in the wrong order or what. Additionally, it really needs some UI improvement for things like when there are options (and recommendations) it should show as multiple-choice (ALWAYS with a way to add an additional note, this annoys me about Claude Code today).
- Why tell the agent you spin up it's not interactive? I really think instead you should hook the AskUserQuestion and proxy that up to Michael and/or the Human directly. I _want_ to see and answer all the questions. Maybe a first pass with an LLM to see if we already have the answer?
- I still feel lost in what's happening and have to dig into agents too often to understand what has been farmed out to them. A LLM-generated summary of their task showing under their name would be nice. The Tasks board is nice but I think it and the "Ask Me" should be global tabs instead of nested under "Michael"
- Triggers seem cool, I don't have a use for them yet but I think this Claw + Agent Harness path is the correct one.
- I keep needing to unstick agents, like Michael, for things asked in the terminal that didn't get put into "Ask Me" or alert me in any way
Semi-related, I'd love some kind of SDK/library to abstract the "control XZY Agent" while being able to expose the raw agent to make writing agent harnesses like this easier. I've been playing with a little agent loop but I often wish I could have it spin up/manage a Claude Code instance and control it (I know it's possible but it's finicky).
This is fantastic. As the little joke I hope it is. Everyone gets their own small disfunctional group, and gets to figure out the challenges of management. You, the manager, are Michael. You know you have to produce something, and you do, but you have no real idea of how. Your diligent agents are Dwight. Overly literal sycophants that are ready to leap to action at your slightest command without any question.
I do think a lot of folk would benefit from the introspection this offers. We've all been given the opportunity to become middle (and middling) managers, and a lot of the challenges we face are those of people who direct. Setting direction is tough. But LLMs are awesome tools.
Insightful. I think Iâd do better at this orchestration at this particular stage of technology development if I named all my agents Dwight, maybe with an occasional Creed.
Not feeling this. Is it necessary to call "agents" by human names? Wouldn't objectives be a safer easier to remember approach to naming? Like, say 'Clips' and 'Seeks'.
I've thought about something like this. When there is a lot of stuff happening at once, it's a mistake to try and communicate it all with text. Agents use tools, reference databases, reference the web, interact with other agents, and spend time processing the information. When you have several agents operating at the same time, communicating what they are doing using some kind of spatial map is a really smart idea.
The office is a decent analog for such a map. Referencing a database? That operation along with processing the information takes a little bit of time. During the interval, have the avatar move to a file cabinet and back to their desk. Have their computer screen change when they access web resources. If they use a particular tool, it can be represented somewhere in the room and used the same way. Interacting with another agent can be similarly represented.
Symbolizing the operation of the agent with movement and behavior is a great way to give an overview. It would support a much richer intuition for how they are accomplishing a task. It wouldn't even need to be a game UI for people who will have a hard time feeling like they are doing serious work while watching what appears to be a game, but that wouldn't bother me.
It is nice to see folks experimenting widely with unique visual representations of agent orchestration, as a consequence of the bottom-up / individual-led development that's driving the industry (inasmuch as it exists yet).
There are folks young enough now that they've never known anything other than "the desktop" (or CLI) as a computing metaphor.
But there's no fundamental reason a metaphor has to be anything specific... it should be whatever is most widely comprehensible and efficient for the problem space.
You are correct this is interesting but it is still quite obviously the wrong UI. For the same reason https://www.gather.town/ ends up being a waste of time.
It's been done before with LLMs. In both cases - human and LLM - trying to "visualize" what people are working on by having a topdown office view is a complete waste of time. It would be like thinking by strolling into a random office IRL and surveying people with their heads down at a random moment you will unlock some great insight.
It's cute. But it's just that. If you want to actually understand some of the things you highlight here there are far more accessible and efficient ways. Like take any data plot or visual graph.
An office is simply the wrong map. It also misses the nature of the joke in The Office. No one is really getting work done. Their work is boring and sucks, no one cares. The whole show is essentially the domino effect of Michael as a rouge agent. They have many scenes where a new character comes in and sees the whole office. It appears normal. Yet after this moment some chaos quickly begins to brew. My point is unless you are planning for a Michael to go haywire, the entire concept of an office is completely uninteresting and tedious. You may as well just use a spreadsheet.
If this dropped just five years ago no one would understand what on earth this software does . In many ways I still don't. Incredible the development we've seen lately, wonder what will stick and what won't?
Why does everyone assume people in the past were idiots? They probably would have figured out what this was for after some critical thinking. Remember those people wrote massive complicated software by hand.
Itâs actually a funny oxymoron if you unpack it - agent implies autonomy and intelligence, harnesses are for unintelligent beasts of burden to control them.
So youâre right, it makes absolutely no sense and would be hard to decipher, but does perfectly capture the contradictions inherent in attempts to treat dumb LLM word generators as intelligent independent agents.
The concept of agents is from the 1970âs, I love how some people think the idea is new. Sure, they didnât have the LLM component, but the rest is the same.
Something can be ridiculous and also conceptually interesting / intelligent.
Gas Town was (transparently, directly from the author's writing) a veneer of visual/linguistic flair applied over a conceptually fascinating and well-thought-out attempt to prod gen 1 LLMs (with all their flaws) into infinite, zero-touch agentic loops.
That the author chose to do so with Mad Max-themed animals doesn't validate or invalidate the underlying concepts.
Personally, I'd rather that than some soulless OpenAI / Anthropic / Microsoft / Google / Meta corporate-scrubbed blob of beige, riskless design.
Especially when it was abundantly clear they realized the presentation choice was ridiculous, but the guts were the more important part.
And not for nothing, there's the 'We're not Kansas anymore' utility of the sufficiently bizarre, to cue people to avoid reusing the wrong prior expectations.
E.g. in Munder Difflin: 'Your LLMs are occasionally brilliant but generally stupid characters, not brilliant human analogs'
an office of your clones. next they'll add a clone HR department to handle the clone performance reviews, and a clone IT guy who's also a clone and keeps filing tickets against himself.
I'm still struggling with setting up long running agents. So far, it's just summoning an agent with a skillset for something specific and then it leaves again. Any tips?
It's interesting how many of these projects are trying to model the worst parts of work (the office, the interaction and information messiness) and then automate that with agents.
With so many AI product launches, I don't even understand anymore what's serious and what's a joke.
Collectively many recent product launches look like we are building a big lab to study sociology using British humour, uttering absurdities with a serious look and then evaluating what sticks.
Not quite verbose but they tend to repeat the same few ideas multiple times with varied wording/imagery. You keep scrolling because you think youâre gonna see something new and by the end youâve realized you just read the same thing 4 times over.
And yet seeing the exact same comment repeated over and over in every thread with an ai generated website is even more verbose. I find it even worse at least the website can contain new information.
Theyâre literally reusing IP from The Office. Itâs clearly not parody, and they directly reference the show. Itâs one thing for a joke project to do that. They are attempting to profit off someone elseâs creative work. And not even in the roundabout way AI does.
Not even in a clever way. The bots don't actually seem to carry any personality traits; they're just ways to label <session_0001> and <session_0002>. I won't accept it as an Office parody if Dwight can work well with others.
Morally, I think it's totally wild to care if some totally unrelated real world thing gets named after a twenty+ year old cultural icon of a TV show as an homage.
They're attempting to profit off their business. The office thing is a theme.
I canât imagine having so little empathy that you canât put yourself in the shoes of the creator of the IP. They didnât even bother to tweak the characters. Itâs not âinspired byâ or riffing on a theme, itâs a straight ripoff being used to market some AI product. Do you think anyone who created The Office would want to be associated with AI, something thatâs threatening to drive them out of a job?
What pisses me off the most is itâs just so damn lazy.
I do like the embrace of The Office as a theme because it so accurately represents the dysfunction of all of the agent swarms Iâve seen: Different personalities pursuing their own little goals that are all competing with each other in subtle ways that eventually leads a funny collapse of the outcome you wanted.
Thereâs no denying that LLMs are getting better by the month, but the current wave of LLM office and personal assistants reminds me of the old trend where people were hiring personal virtual assistants from foreign countries to manage their work and interface with people. It seemed like an obvious time saver but every time I interacted with someoneâs virtual assistant it felt like I was playing a little game of navigating hidden structures and communication barriers to get the message I needed to the person hiding behind it all. Then there were the inevitable scheduling failings, missed meetings, dropped emails, and other things that get blamed on the assistant. Iâm getting deja vu watching it all happen again with LLMs replacing virtual assistants and outsourced teams.
Maybe these concepts work for people who are trying to solo dev and who find it interesting to set up and debug little systems for everything they do, but I really donât like working with anyone who surrounds themselves with one of these multi-agent coordination systems as an external shell. Keep it to the internal work and maybe itâs fun for some people.
In areas of applicability where GenAI to now considered to be good/great, the pattern I've observed is a frighteningly rapid progression from "abjectly terrible", to "middling and mostly more trouble than it's worth" and then "good/great".
Agentic coding and image generation are probably the best examples of this.
There's no guarantee that this trend will be seen with virtual assistants, but if it does, we're on the brink of incredible. Dedicated personal assistants, that actually know and understand one's context and work well together, are unarguably hugely appreciated by those who can afford them.
Hey guys, thanks for putting it here, I am Chaitanya I built Munder Difflin, I am here to answer all your questions(except nylonstrung).
For people who haven't tried it: Munder Difflin is a local multi-agent harness that wraps around your existing claude code and codex subscriptions(we literally support almost all harnesses/coding agents).
Simulations are deterministic, they do not consume tokens, infact most of the users(20K+ in a week) say that it has reduced their token consumption due to a benchmarked memory layer acting as a hive mind called mempalace.
Common use cases apart from coding: 1. Create triggers that runs an live agent with your context(Webhooks, slack, scheduled) 2. Almost any kind of automation for yourself(I make it review PRs, send cold emails with enriched context, manage discord, Send myself analytics about how app is doing on email an end to end AI video production and posting workflow in 1 prompt and then some)
I'd love to hear your feedbacks on it.
It's interesting but It was hard to tell from a quick read if this was a fun game with LLMs or a productivity tool. You could stand to make that more clear.
...I agree but I also kind of like it? More software should aim to be both useful and fun.
SO WEIRD.
My flagged comment has been unflagged, but the comment I made referring to the odd flagging is collapsed but not actually flagged either.
I think I must be in some kind of moderation hell.
EDIT: Oh - apparently someone at HN did it. But that's a shame, because there was good content in that thread. Why are you guys hiding it?
my last comment was flagged - presumably i didn't make it pedantic enough for someone, so let me try again:
why do you feel the need to pigeon-hole this work? the greatest things in life are more than one thing at once.
EDIT: stealth edits of your comments is naughty-naughty.
I donât think GP is being unreasonable. Half of this webpage kinda sounds like parody and the other half serious. Itâs valid feedback.
did i say he was being unreasonable?
i asked him why he's pigeon-holing it. and you, too? don't hackers enjoy a little bit of playful ambiguity in their lives? must it be all serious all the time?
sheesh.
Sir, this is the Internet in 2026. We don't allow playful anything unless the intent is clearly labeled and approved by your local decency csar. Please continue to refrain from double entendre, sarcasm, irony, or any other form of misunderstandable communication. Every thought must be of clear intent to the community so we may apply the correct direction of moderation to facilitate the amplification or suppression of thoughts in line with the already accepted and centrally approved community morality. Thank you for your cooperation.
UNDERSTOOD. WILL COMPLY.
Ya damn right u will ;P
porque no los dos?
When you say no code leaves your machine, how does code get evaluated by your LLM of choice? Doesn't Claude see your code when you send a prompt just like it usually does?
âsend cold emails with enriched contextâ
Whatâs the difference between cold and spam?
Nailed it. I constantly get âcold emailsâ that were clearly written by an AI that read 2-3 pages on my web site, my LinkedIn bio, and something else on my social media. They all follow the same formula: try to break they ice with something funny about somewhere I took a photo recently, then segue into what I do for work, then offer to help me do it better - usually by offering leads or CRM improvements.
Itâs such an obviously repetitive recipe that someone must have published as an example somewhere, and now lots of people (or bots) are trying to follow it, all mining for gold that doesnât exist.
I have no desire to get either so functionally nothing.
most cold emails end up spam, only a small fraction end up being useful. My cold email strategy respects google's limits are very personal and mostly not for selling.
Why I mentioned this here: Everyone knows about cold emailing, this is one of the ways I describe the personalisation and integration. It does, slack, discord, and practically works with anything that has an api key.
I often get cold emails from people interested in my very specific hobby (see website in profile). These are desired and fun.
Mind, they are not written by LLMs. I would consider that spam, outside of very limited circumstances (maybe a non-english speaker using an LLM for translation).
The name.
Cool project! Iâm still trying to figure out if this sort of thing is the future, or something that the big AI labs will subsume into their products on their own soon. Or something that wonât be necessary in a few months when LLMs get smarter.
You won't see big AI companies allowing other providers in their native agents/harnesses. I believe this is a great opportunity for open source projects like us.
I don't want to contact you for pricing.
Too bad, man's gotta make money to live in the capitalistic society.
That is possible while openly stating your products price.
Hey thanks, I'll update the pricing on website today.
Does it work with local LLMs?
Yes, it supports opencode and qwen clis, both of them can be configured to work with local LLMs.
when did you launch, how many users you have, and what's the business metrics if you don't mind sharing.
We launched 8 days ago on product hunt, been working on it for 2 months before that.
We are piloting our TEAMS plan with 5 businesses and currently have over 10,000 end users and over 30,000 agents spawned on the community version.
About to launch our PRO tier for the power users of community version. We are optimising for everything that'll increase their productivity.
Hope it helps, also I am looking for an angel investor if anyone's interested in making this big.
How is this different from a Hermes agent? Any major pro? Besides the cute graphics.
And why do you need E2E encrypted comms between your own agents? What is that protecting against? To prevent other agents snooping and maybe getting derailed? Or do you support conversations between agents of different users?
Let's say if you are away from your computer and your collegue asks you for a report that's anyways gonna take you one prompt on your agent.
Your Collegue's Munder Difflin agent can directly talk to your Munder Difflin agent through our network, you'd want your messages end to end encrypted.
What an exciting project, Chai. Congratulations on your work.
Thankyou so much <3
don't let the crankles get you down. this thing is delightful.
This is still better than what I got at some of the subreddits, I can handle it.
HEH. It still doesn't make it right.
I love it!!! But please, change that "GOD orchestrator" label, as a believer, I find it profane and offensive.
To get past the uh⌠man I really lack the words to describe the responses that youâve already received, but I guess to avoid them going forward, I would re-label the point your making as being more culturally sensitive. Itâs not asking people to accommodate your faith, rather your culture. It removes the convicting tone and imposition of your beliefs that others wish to derive from your request and instead communicates a desire for representation that isnât âmandatedâ but appreciated.
It probably wonât prevent people from getting worked up over the request but it sheds a lot of potential hate and ammunition.
Perhaps consider reframing the term from referring to the God in which you believe but as merely a reference to just the top agent in its little Munder Difflin universe?
Seems like both a mental model which more accurately represents the meaning of the "GOD orchestrator" term, and solves the meaning disconnect closest to the source that is troubling you, thus being the most effective and efficient solution for everyone?
Everyone is a believer.
What exactly do you believe and why should anyone accomodate it?
Is that satire?
Ok, I've been running it for a couple hours and below are my thoughts. Please note that I do find it fascinating even if most of what I'm about to say is complaining about the parts I like less.
- Pipelines, not agents. Roles, not agents. I really don't like the idea of defined agents with their own prompt. I want to define roles and spin up N agents with that role. Furthermore I want pipelines "Plan -> Review Plan -> Approval Gate -> Develop -> Code Review + Fix loop -> QA -> Approval Gate -> Merge -> [Ship]". I don't like the work just bouncing around seemingly randomly
- Settings don't seem to save/persist? Or some of them don't. I couldn't let "Michael" spin up agents on "his" own and then randomly he did it even though the setting was still off. Settings has the normal LLM jank I've seen.
- macOS Notifications are broken, they send for any little reason, and then they don't send when you're actually needed. It's like each agent finishing a round causes a notification.
- Speaking of missing notifications, the _most important_ screen to me is the "Ask Me" tab under "Michael", where they ask questions (more on that later) but there is zero indication that anything is waiting for you. You have to dig into it yourself.
- The "Ask Me" tab is great.... when it works. I've had to unstick agents or answer questions they were waiting on answers for
- Trying to be too cute, it was cute for a minute, now I don't care (and I _love_ The Office). I want a more utilitarian view. I want to see questions, plans, be able to inject new ideas, and a small overview of what each agent is doing. I don't need half the screen taken up with a "game ui".
- Why no clear? I don't understand at all the idea of them keeping context. Maybe I'm missing something and I shouldn't be using persistent agents except for more persistent jobs (like Michael's?).
It's an interesting concept, very "Gas Town", and it make me want to write my own that does more of what I'm looking for but I don't have the time (or tokens) currently to take on another project. My current best approach of herdr+6-10 Claude Code sessions feels like it works better than this and keeps me close enough to the decisions I want to make.
I think you and I are cut from the same cloth. Role-based pipelines and coordination with scoped permissions generated at runtime and the ability to communicate is the end game in my opinion, for now at least. Orchestration finally clicked with me when I tried out Gas Town six months ago and I immediately created a system to offload it to the cloud by using a combination of GCP CloudRun + Firestore (for temp file storage and to fill role of beads). Right when I finished ironing out the kinks Anthropic laid down the hammer on how users can use subscriptions and I abandoned it.
Now I am using custom pipeline/messaging control plane that spawns microvms with scoped creds minted at runtime. I had to make my own control plane specifically because everyone is focused on agents with personalities rather than roles and pods of workers. I like everything about Buzz except for that.
I'm slowly inching towards my own control plane because my skill-based workflows (not Claude Code Workflows) are brittle but still work better than anything else I've tried. Enforcing the pipeline has been mostly successful but I'd greatly prefer something more structured.
I've been dipping my toes into the waters of higher-level orchestration for a while now but keep scurrying back to driving CC mostly directly. Any extra layers I've added have just caused too much waste (time & tokens) or otherwise produced inconsistent results. I want to "live" in the decision making space. I want to be consulted for decisions (with context provided so I can context-switch easier), I want to review plans, I want to review the actual UI (when there is any), and I want to be able to easily inject new ideas/tickets into the machine (and probably so way to adjust issue priority).
Thankyou so much for the feedbacks, I'll pick them up in next few versions. If you try it again in a week or so, you'd find a lot of these issues resolved.
Some more observations
- It seems like letting Michael spin up things is the prefered pathway instead of defining a bunch of agents.
- The Ask Me really needs some love. The "X/N" display doesn't make sense to me, I started on "5/5" and as I answered the questions it went to "4/5" -> .... "1/5", I'm not sure if the questions were asked in the wrong order or what. Additionally, it really needs some UI improvement for things like when there are options (and recommendations) it should show as multiple-choice (ALWAYS with a way to add an additional note, this annoys me about Claude Code today).
- Why tell the agent you spin up it's not interactive? I really think instead you should hook the AskUserQuestion and proxy that up to Michael and/or the Human directly. I _want_ to see and answer all the questions. Maybe a first pass with an LLM to see if we already have the answer?
- I still feel lost in what's happening and have to dig into agents too often to understand what has been farmed out to them. A LLM-generated summary of their task showing under their name would be nice. The Tasks board is nice but I think it and the "Ask Me" should be global tabs instead of nested under "Michael"
- Triggers seem cool, I don't have a use for them yet but I think this Claw + Agent Harness path is the correct one.
- I keep needing to unstick agents, like Michael, for things asked in the terminal that didn't get put into "Ask Me" or alert me in any way
Semi-related, I'd love some kind of SDK/library to abstract the "control XZY Agent" while being able to expose the raw agent to make writing agent harnesses like this easier. I've been playing with a little agent loop but I often wish I could have it spin up/manage a Claude Code instance and control it (I know it's possible but it's finicky).
Pipelines and roles are a good point.
This is fantastic. As the little joke I hope it is. Everyone gets their own small disfunctional group, and gets to figure out the challenges of management. You, the manager, are Michael. You know you have to produce something, and you do, but you have no real idea of how. Your diligent agents are Dwight. Overly literal sycophants that are ready to leap to action at your slightest command without any question.
I do think a lot of folk would benefit from the introspection this offers. We've all been given the opportunity to become middle (and middling) managers, and a lot of the challenges we face are those of people who direct. Setting direction is tough. But LLMs are awesome tools.
Insightful. I think Iâd do better at this orchestration at this particular stage of technology development if I named all my agents Dwight, maybe with an occasional Creed.
Not feeling this. Is it necessary to call "agents" by human names? Wouldn't objectives be a safer easier to remember approach to naming? Like, say 'Clips' and 'Seeks'.
Safer? From a mental health perspective? Tbh I like human names
um no. from a technical design perspective, obviously.
I've thought about something like this. When there is a lot of stuff happening at once, it's a mistake to try and communicate it all with text. Agents use tools, reference databases, reference the web, interact with other agents, and spend time processing the information. When you have several agents operating at the same time, communicating what they are doing using some kind of spatial map is a really smart idea.
The office is a decent analog for such a map. Referencing a database? That operation along with processing the information takes a little bit of time. During the interval, have the avatar move to a file cabinet and back to their desk. Have their computer screen change when they access web resources. If they use a particular tool, it can be represented somewhere in the room and used the same way. Interacting with another agent can be similarly represented.
Symbolizing the operation of the agent with movement and behavior is a great way to give an overview. It would support a much richer intuition for how they are accomplishing a task. It wouldn't even need to be a game UI for people who will have a hard time feeling like they are doing serious work while watching what appears to be a game, but that wouldn't bother me.
It is nice to see folks experimenting widely with unique visual representations of agent orchestration, as a consequence of the bottom-up / individual-led development that's driving the industry (inasmuch as it exists yet).
There are folks young enough now that they've never known anything other than "the desktop" (or CLI) as a computing metaphor.
But there's no fundamental reason a metaphor has to be anything specific... it should be whatever is most widely comprehensible and efficient for the problem space.
LLMs remade Clippy. This sounds like using Microsoft Bob to manage them.
Imagine how much more infuriated people would have been if Clippy wasn't a cartoonish-looking paperclip, but a human facsimile.
Sometimes absurdity or comedy is the point when you're developing a known-partial solution.
You are correct this is interesting but it is still quite obviously the wrong UI. For the same reason https://www.gather.town/ ends up being a waste of time.
It's been done before with LLMs. In both cases - human and LLM - trying to "visualize" what people are working on by having a topdown office view is a complete waste of time. It would be like thinking by strolling into a random office IRL and surveying people with their heads down at a random moment you will unlock some great insight.
It's cute. But it's just that. If you want to actually understand some of the things you highlight here there are far more accessible and efficient ways. Like take any data plot or visual graph.
An office is simply the wrong map. It also misses the nature of the joke in The Office. No one is really getting work done. Their work is boring and sucks, no one cares. The whole show is essentially the domino effect of Michael as a rouge agent. They have many scenes where a new character comes in and sees the whole office. It appears normal. Yet after this moment some chaos quickly begins to brew. My point is unless you are planning for a Michael to go haywire, the entire concept of an office is completely uninteresting and tedious. You may as well just use a spreadsheet.
If this dropped just five years ago no one would understand what on earth this software does . In many ways I still don't. Incredible the development we've seen lately, wonder what will stick and what won't?
Discovery is part of good design. It is easy to confuse complexity with sophistication.
Why does everyone assume people in the past were idiots? They probably would have figured out what this was for after some critical thinking. Remember those people wrote massive complicated software by hand.
I didn't imply people in the past were idiots, but how on earth would anyone know what an _agent harness_ is, say?
Itâs actually a funny oxymoron if you unpack it - agent implies autonomy and intelligence, harnesses are for unintelligent beasts of burden to control them.
So youâre right, it makes absolutely no sense and would be hard to decipher, but does perfectly capture the contradictions inherent in attempts to treat dumb LLM word generators as intelligent independent agents.
I don't think it's an oxymoron at all. An expert climber still uses a harness for safety, even though they won't (typically) try to jump off the wall.
They would 1 times out of 10 if they were an LLM. Then say sorry, youâre absolutely right I shouldnât have jumped off the cliff.
From beyond the grave
AlphaGo already had a harness. An AI in a loop with tools is not exactly a Black Swan.
Because the term or concept of âAI Agentâ existed long before LLMs??? JesusâŚ
The concept of agents is from the 1970âs, I love how some people think the idea is new. Sure, they didnât have the LLM component, but the rest is the same.
You're addicted to being mad
This is super cute. I haven't tried it or anything but it's really fun, seems genuinely useful too. I don't really get people calling it cringe.
They're thinking it's serious like Gas Town, that's why
Something can be ridiculous and also conceptually interesting / intelligent.
Gas Town was (transparently, directly from the author's writing) a veneer of visual/linguistic flair applied over a conceptually fascinating and well-thought-out attempt to prod gen 1 LLMs (with all their flaws) into infinite, zero-touch agentic loops.
That the author chose to do so with Mad Max-themed animals doesn't validate or invalidate the underlying concepts.
Personally, I'd rather that than some soulless OpenAI / Anthropic / Microsoft / Google / Meta corporate-scrubbed blob of beige, riskless design.
Especially when it was abundantly clear they realized the presentation choice was ridiculous, but the guts were the more important part.
And not for nothing, there's the 'We're not Kansas anymore' utility of the sufficiently bizarre, to cue people to avoid reusing the wrong prior expectations.
E.g. in Munder Difflin: 'Your LLMs are occasionally brilliant but generally stupid characters, not brilliant human analogs'
>They're thinking it's serious like Gas Town, that's why
It's not? https://munderdiffl.in/#pricing
I guess I wouldn't use it, but I just bought two of the asset packs of the pixel artist instead.
an office of your clones. next they'll add a clone HR department to handle the clone performance reviews, and a clone IT guy who's also a clone and keeps filing tickets against himself.
Can't be worse then the preformative just-have-to-be-done newage bs HR reviews in modern big corpo landscape.
I'm still struggling with setting up long running agents. So far, it's just summoning an agent with a skillset for something specific and then it leaves again. Any tips?
Would running this in say a KASM workspace fall within the open source license?
It's interesting how many of these projects are trying to model the worst parts of work (the office, the interaction and information messiness) and then automate that with agents.
With so many AI product launches, I don't even understand anymore what's serious and what's a joke.
Collectively many recent product launches look like we are building a big lab to study sociology using British humour, uttering absurdities with a serious look and then evaluating what sticks.
hey neat stuff. I'd actually like the option for the simulation NOT to be deterministic. I don't mind sacrificing 100% of any productivity for this.
1. Pranking: Rejected my PR? I'll put your keyboard in jello. Then the jello'd agent's prompts all get your keyboard is currently in jello attached.
2. Office Romance: Certain agents prefer to work with each other, but can randomly experience entertaining breakups.
3. Dundies style award ceremonies. The titles can show up under their names until the next ceremony i.e. "Hottest Agent in Office"
Reference to the office is hilarious, this might actually get as much work done as Michael did in the show
This is funny and shamelessly bad at the same time. I have no idea how it's getting so much attention.
I don't understand what is this used for. Is my kitty terminal not enough of a multi-agent orchestrator? Genuinely lost.
This looks like a tool from Cowboy Bebop.
Is it just me or do AI made websites tend to be too verbose?
Not quite verbose but they tend to repeat the same few ideas multiple times with varied wording/imagery. You keep scrolling because you think youâre gonna see something new and by the end youâve realized you just read the same thing 4 times over.
Hasn't that been true of marketing websites for a long time?
LLMs in general have a tendency to be overly verbose, in my experience.
And yet seeing the exact same comment repeated over and over in every thread with an ai generated website is even more verbose. I find it even worse at least the website can contain new information.
Please pardon. I have seen people complaining that copy being written with an LLM but did not see any mention of verbosity.
This is cute and fucked up at the same time.
GOD Orchestrator lmao
This project is cringe and I really hope we see less stuff like this
People who who "cringe" as an adjective are cringeworthy, and I hope we hear less from them.
Calm down Angela.
I wouldn't call it cringe but I don't think it adds a lot of value, distracts me even more & even pulls unnecessary resources I'd guess?
but if people need some tamagochis to look at while their AI is working, maybe it makes sense
The next steps in innovation will occur at the harness level. Why not have some fun while weâre at it?
That you, Stanley?
Why?
I hope we see fewer comments like this that aren't shitting on projects that you have no stake in anyway
This is a business--they have cloud and a sales team.
Fair. I'll rm the personal project bit
Eh, so what? It's whimsical and fun, and it's better if more things like this exist in the world than not.
Theyâre literally reusing IP from The Office. Itâs clearly not parody, and they directly reference the show. Itâs one thing for a joke project to do that. They are attempting to profit off someone elseâs creative work. And not even in the roundabout way AI does.
Not even in a clever way. The bots don't actually seem to carry any personality traits; they're just ways to label <session_0001> and <session_0002>. I won't accept it as an Office parody if Dwight can work well with others.
Legally, I agree.
Morally, I think it's totally wild to care if some totally unrelated real world thing gets named after a twenty+ year old cultural icon of a TV show as an homage.
They're attempting to profit off their business. The office thing is a theme.
I canât imagine having so little empathy that you canât put yourself in the shoes of the creator of the IP. They didnât even bother to tweak the characters. Itâs not âinspired byâ or riffing on a theme, itâs a straight ripoff being used to market some AI product. Do you think anyone who created The Office would want to be associated with AI, something thatâs threatening to drive them out of a job?
What pisses me off the most is itâs just so damn lazy.
Honestly, making Ricky Gervais upset might be the most redeeming thing in this product.