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.
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?
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?
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 I’d 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'.
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'
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.
The concept of agents is from the 1970’s, I love how some people think the idea is new. Sure, they didn’t have the LLM component, but the rest is the same.
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.
They’re literally reusing IP from The Office. It’s clearly not parody, and they directly reference the show. It’s one thing for a joke project to do that. They are attempting to profit off someone else’s creative work. And not even in the roundabout way AI does.
Not quite verbose but they tend to repeat the same few ideas multiple times with varied wording/imagery. You keep scrolling because you think you’re gonna see something new and by the end you’ve realized you just read the same thing 4 times over.
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.
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.
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.
why do you feel the need to pigeon-hole this work? the greatest things in life are more than one thing at once.
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.
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?
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.
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'
It's not? https://munderdiffl.in/#pricing
but if people need some tamagochis to look at while their AI is working, maybe it makes sense