There’s 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 someone’s 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. I’m 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 don’t 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 it’s fun for some people.
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.
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.
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?
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 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.
What’s the difference between cold and spam?
It’s 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 doesn’t exist.
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.
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).
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.
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?
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.
It probably won’t prevent people from getting worked up over the request but it sheds a lot of potential hate and ammunition.
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?
What exactly do you believe and why should anyone accomodate it?
- 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.
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'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).
- 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).
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.
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.
Sometimes absurdity or comedy is the point when you're developing a known-partial solution.
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.
So you’re 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.
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
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.
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"
but if people need some tamagochis to look at while their AI is working, maybe it makes sense
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.
What pisses me off the most is it’s just so damn lazy.