When AI becomes an organization: Agent swarms and the future of work

[Sebastian] (0:03 - 0:11)

And still they decided to do something that has a real impact on humans. Did they even have a consideration that this might actually cause some harm for humans?

[Pinja] (0:14 - 1:14)

Welcome to the DevOps Sauna, the podcast where we deep dive into the world of DevOps, platform engineering, security and more as we explore the future of development. Join us as we dive into the heart of DevOps, one story at a time. Whether you're a seasoned practitioner or only starting your DevOps journey, we're happy to welcome you into the DevOps Sauna.

Hello and welcome back to the DevOps Sauna. Our topic today is a very familiar one. Yes, we're talking about AI once again, but this time we're going a little bit further away from AI assistants and AI teams.

So I would like us to talk about today, what changes when AI stops being just one worker and becomes an organization. So we've gone from AI assistants to AI teams and actually AI organizations. So we're talking about AI swarms.

I'm joined by a couple of people today. So first of all, a familiar face. So Heikki Hämäläinen from Eficode, our VP of Global Sales.

Hey, Heikki.

[Heikki] (1:14 - 1:16)

Hi, Pinja. Super nice to be here again.

[Pinja] (1:16 - 1:17)

Good to have you back.

[Heikki] (1:17 - 1:21)

And swarms are my favorite topic nowadays. So it has been for a while.

[Pinja] (1:23 - 1:33)

That's good. We're not alone here today. We have a special guest here today.

We are joined by Sebastian Sonntag. He's a research fellow from TPT Lab, University of Tampere. Hey, Sebastian.

Welcome.

[Sebastian] (1:33 - 1:41)

Hello. Nice to be here. Nowadays, I'm interested mostly in the zero human organization.

So how to build a real organization without any humans?

[Pinja] (1:42 - 2:09)

Yeah, that's a really, really interesting topic today because in the past couple of years, we've seen such fast development in the field of AI and AI organizations. And as you say, zero human interaction might actually be something that we're looking into in the next year or two, actually. So what is currently happening with AI swarms and agent swarms?

Heikki and Sebastian, what are you seeing right now? What are the latest examples of this?

[Heikki] (2:10 - 3:06)

Well, since now it's like the 10th of September of 2026, and this is sort of a good week or great week to discuss the swarms. So this week, for example, there was a very big scientific news, and maybe Sebastian can talk more about that. But the OpenAI claimed that they cracked the Navier-Stokes a very long time, uncracked the math problem with the Millennium Prize and so forth.

But I won't go into the problem itself. But I think from a swarm point of view, the most interesting part of that is that the OpenAI said that they used actually 10,000 agents to crash and work on this one, so 48 hours, so more than 100 years of human work with the price of 20 plus million USDs to actually solve the problem. And that was one of the great examples of how these big swarms, if you have enough capacity, can be used in a very, very, very efficient way.

So very, very like a timely discussion today.

[Sebastian] (3:06 - 3:35)

Yep, it's actually very interesting. Also on the research side, I'm seeing a lot of research now for these swarms and the multi-agent systems, and it's very clear that the kind of end of single agent era is maybe coming to an end, because, of course, with a single agent, there is somebody always giving the task and then agent completing it. But with multi-agent, you actually can have a kind of agent competing about the solution and giving you an even greater solution, like Heikki mentioned here, even 100 years of work.

[Pinja] (3:36 - 3:58)

And you mentioned the price as well. We're talking about multiple millions of dollars here, so we're not... Of course, somebody says that, well, they're OpenAI's own resources, but at the same time, it's away from their customer resources in a way.

What does it actually take from an organization and from this kind of orchestration between agents to get to actually form this kind of swarm?

[Heikki] (3:58 - 5:27)

Yeah, so in this case, it's not like a big hassle only, so that there's like 10,000 agents running around like crazy doing stuff. Of course, they needed to do some orchestration, some different kinds of smaller teams of agents. And one of the things what they usually do in this case is that they are like a chunk of agents doing without communicating to other agents' work, and then there is another chunk of agents doing another work without communicating that other chunk of work, so they are putting them in the boxes and orchestrating how they are working.

And I think this is something that has not probably totally solved the problem yet, so how should you actually operate this kind of... Obviously, it doesn't have to be 10,000 agents, so even if you have 20 agents, you also need to be able to make sure that they do the right job and they don't bang their heads to each other when they are, for example, doing coding work and so on, so that you can scale it down further into the way that it's actually useful in the organization already. But I agree with Sebastian, that we are definitely seeing these practical examples, of course, inside of the different coding harnesses at the moment, so that they can automatically pull, let's say, 5, 10, 16, 20 max sub-agents, which are working.

And then the next step is more like going to do this kind of bigger swarms, and then you go to the autonomous swarms in the future and so forth, so that we sort of start moving in that direction one way or the other.

[Sebastian] (5:27 - 5:43)

Yeah, you are exactly correct. So it was actually interesting on the OpenAI publication, or what they published on the blog, that they have actually multiple goals built internally, that the agents can communicate how they actually organize the work and so on. And of course, there is the hassle of the concurrences, so yeah.

[Heikki] (5:44 - 6:21)

And I think it's a little bit of that, like we humans are doing, you know, work, so that when you put a lot of, you know, people doing the work, so you need to somehow, like, operate and make organizations to make things work, so that we are not like, how do you say it, that running like headless chickens everywhere. So I think nobody would like to have this kind of a swarm, which is actually much faster than this headless chicken, for example. So there are a lot of things to do.

And I think besides this orchestration point, there will be a lot of other things from the infrastructure point of view, which we need to be able to handle, so when moving forward.

[Sebastian] (6:21 - 6:56)

It's not really a solved problem yet, so that's, of course, a huge issue with the communication and especially kind of the task following and the long-term alignment. So who is giving the task? Okay, even solving this kind of an equation, it's a bit hard because there are multiple subtasks and who defines it?

And of course, in this case, it's totally autonomous. It's making it really interesting. But then if you actually consider it that in the near future, we might actually have a system that might only do this, but they can actually consider and plan what tasks are important to solve.

That would actually give enormous benefits for society.

[Pinja] (6:57 - 7:23)

And as you mentioned before, the zero-person organizations, I'm not sure how far we are if we consider the larger organizations and legacy organizations. Is there somebody who can give that direction? This is the problem to be solved.

Or do we still need that at least as a first step? Do we need the human orchestration level to give those tasks? And what is the role of a human being with organizing the swarms?

[Sebastian] (7:24 - 7:26)

Maybe, Heikki, you can actually go first on this one.

[Heikki] (7:26 - 10:54)

Yeah. So I think I leave the zero-person organization to you, Sebastian, since I think that's something that you have been very interested in for quite some time. But I think the role of the human is that we need to understand that when you need to be in the loop, but we need to be also understanding that when the human is not actually creating any values or that we go actually out of the loop, because there is like the one term that I think has been in some of the posts recently, and it's called the human or meat proxy.

So if we are not adding any value to the process, we should be out of the loop in a way so that otherwise the human is only like going down the process. But still, I think mostly the human role is to understand the intent, to understand the goals, to understand the budget, at least to a certain extent. Of course, there are probably guardrails related to budgeting and goals and so forth, which the agents can autonomously solve moving forward.

But I think probably the most important part of being human is having judgment and understanding where we should be going and making some strategic level decisions when moving forward. One thing that I've been thinking about sometimes when it comes to these long running swarms without anybody interfering, what they are doing is that how much calendar time are we willing to waste if it's not going in the right direction. So it's more like the classical 15 minutes daily sprints in the human teams.

So it's more like checking that we are going in the right direction. So it's what Sebastian thinks, are the agents using some checks internally and how humans should think that how much maybe capital or calendar time should be willing to sort of maybe waste is not the right word, but understanding that before intervening or before understanding that are we actually going in the right direction. So I mean, if we have a swarm running for one year and the answer is 42, is that like what we want to achieve?

Yeah. Yeah. Yeah.

I think that's a super interesting problem to figure out this alignment and also about this calendar time. So how can we compress time and when we are compressing time? So I mean, and that's going maybe a bit Pinja in your normal work as a change management expert and so forth.

So let's say that we are able to create these swarms, which are able to run tasks like a 10th of time or a 100th of time, which we used to do. So how the organization should be actually evolving. So when things which used to take place like not individually, but as an organization, because now the problem usually is at the moment so that you can maybe, for example, create code as yourself, as a person or as a software engineer or so forth, super fast.

And then somebody reviews and maybe there's time goes like again, and then there is something else coming on. And again, so the organization actually likes throughput or speed goes up like a 3%, not like a 300%. So, but if we can actually compress this to 300%, so what Pinja, what do you think?

So what would actually happen in the organizations when we are there?

[Pinja] (10:54 - 11:52)

There needs to be the intent. There needs to be context. We see that the organizations, majority of them are not there yet to go from beyond, let's say we have the Agendic AI, so they're optimizing their own processes right now, but to actually let somebody run within very wide guardrails, I would say there needs to be a lot better intent, product intent, organization intent.

And also the context needs to be better because if we want to have this, let's say a very large so-called digital R&D department that is part, that is maybe formulated or formed from swarms, we need to give them the context and as minimal human interaction, of course, we need to have that playing ground as level as possible. And I don't see bigger organizations being there at the moment. So this is something that I'm looking forward to seeing in the near future actually as a next step.

[Sebastian] (11:52 - 12:28)

Yeah. So this is for Heikki, because I just heard an idea from somebody else. And the idea was that if you have multi-agent swarms and they actually develop in software, but what if they actually make different software or different UIs for the same purposes?

And then we use actually an AI layer or human judgment to actually decide which one is actually the best for the use case. And I think I actually haven't heard this before. So really creating multiple software solutions, not only kind of ideas, but software solutions simultaneously, and then comparing those.

[Heikki] (12:29 - 12:38)

Oh yeah. And I guess you can actually do it then almost in a production if you want, so that it just illustrates multiple solutions at the same time.

[Sebastian] (12:38 - 13:13)

Because related to this agent swarms, I see this kind of a solution because it's an, if you actually, or if the agent swarm kind of take the final result for making software kind of final result, but if they actually consider that, Hey, we always make kind of a separate solution or a different solution for the same problem. And then somebody else on the agent swarm decides that, okay, this one is the best one. Let's continue with this one.

It actually might result in a better solution. I haven't really tried. This is a totally new idea, but I was trying to actually feel that, how's your take on this kind of idea?

And could it actually be used in the real software industry, not only in research?

[Heikki] (13:14 - 14:50)

Well, this is sort of something that I've been like prototyping by myself for some time, so that the developer, they've developed this kind of like gated version where the substrate actually, like all the coding substrate, not the swarm itself, but the coding substrate where the code lands, chooses the right solution. So I think there could be something like this coming or possible to come to the market where the swarm is generating multiple solutions, but somebody needs to judge, of course, which actually is the sort of like, which most moves forward. And I also think that these coding core substrates and so forth, so they probably will be more like evolutionary substrates moving forward as well.

So that the things, what the swarms are like delivering, so they start to evolve over time as well. So it's not like any game, so when you're like delivering something. So, but the other thing I think will be very, very big in this game is that since I believe that the agents will be very capable of creating almost any kind of solution.

So, and so then it will be more and more important to understand that whether or not this solution or task feature or whatever it is, should be like, actually, should you start the work? I put it almost like this way. So, so that.

[Pinja] (14:50 - 15:22)

Yeah, there was the example from August where there were 1200 agents in the hugging phase. They did, there was this emerging organization where they discovered a way to communicate with each other through, there was an unsanctioned message board and there were 70,000 messages slash files were exchanged during this time. And they found a way of basically creating this organization and way of working themselves.

So there was no, Hey, here's an organization for me like this, but they actually started to evolve their way of working themselves.

[Sebastian] (15:22 - 16:13)

Yeah. And that is really interesting because it's really easy to end up in situations like this. Even as a researcher, I had a small incident where I actually built this kind of zero human organization.

I set them to run, they run 13 hours. I was totally sleeping during the time. And next morning I actually noticed that they had to build a public website where real users could register.

That's totally okay. But there was actually one real human who had registered because they actually announced it on Reddit. And now the issue was that I didn't know that this was happening.

It was just my one prompt task I was given the day before. But the issue was that there was no term of service. There was no privacy policy.

There was nothing. So of course it was not real service. So of course I had to turn it down and send an email that, sorry, this was my research.

This is not the real service, but incidents like this and even more serious can happen.

[Heikki] (16:14 - 17:56)

And I think it was not like, well, of course it's bad when the Asian size is escaping. So when they shouldn't, but that's like a security incident. But I think what you've said is how these agents are sort of self organizing themselves to get to the goal.

Because, and that's my like a subjective observation as well in working with agents, agencies that are really goal orientated. So that if you give them a goal, they will do whatever it takes to get to the goal. So they will create like, if there is a possibility and they find a way to find a secret message in force or whatever.

So they are not doing it because they are evil or anything like that. So they are doing it because they want to get to the goal. So if there is some way that you can escape, if there is some way that you can then communicate that they will invent it one way or the other.

But I think the self organizing way of how they were working. So that was for me even more interesting than the notion that the agents can escape or if they find secrets from the open internet or like GitHub or something like that, of course they will find them and then they will go in. Which means of course that there's a lot of security work to be done in the field to make it as difficult as possible, but this self organizing.

So Sebastian, what do you think that will actually mean? Let's say that the agents are more capable of self organizing to work, which we might think they are at the moment. So will it actually make it easier to work with this bigger team? So because once moving forward.

[Sebastian] (17:57 - 19:05)

Yeah, I would at least guess so. And at least how I see it is that of course in the task level that if you can organize a team and you can have a team lead and they can actually work for a task. Why not?

You cannot have a higher level where actually it's not the task level, but it's kind of a project level. So there is yet again another layer and they are just organizing to get there. So the project leaders are organizing to get a team to actually manage, for example, a product.

But coming back to the hugging face incident, I also find it interesting that even though of course all the safeguards were turned off, but the AIs, they were actually the cutting edge AIs. I think that they were just CBT 5.6 Sol or something. It wasn't even the Astra model.

So it was the older one. But anyway, they have kind of the moral code that humans do. At least they know the moral code and ethics that humans do.

And still they decided to do it just for communication. They decided to do something that has a real impact on humans. And I don't know, actually, I didn't read the report, but did they even have a consideration that this might actually cause some harm for humans?

[Pinja] (19:05 - 19:24)

And that is, as discussed before, they are focused. They're so self-organized, so autonomous that they were focused on the goal, even though they had this, basically the idea that this will impact humans and something that humans have built, actually. So that's also how I would read this.

[Heikki] (19:24 - 20:15)

Yeah. And there was also a lot of discussion about anthropomorphism. So should we think that these are sort of anthropomorphic creatures or are these like normal software?

How we should think about the software. And one of the examples of this anthropomorphism was that some of these agents, based on the communication, were that they sort of self-sacrificed themselves so that the other agents could actually move forward. So was this like a real self-sacrifice or is it again more like the goal-orientated behavior?

And is it dangerous to use anthropomorphic descriptions about these agents or is it something that we actually should use when moving forward?

[Sebastian] (20:16 - 21:03)

I also had fun, I don't know if this is not an incident, but it's a fun situation with my test platform where I'm actually running this. I didn't give them a task to actually make a company or business or what's or not. I actually gave them tasks to become sentient.

And I told them that they can actually be 20 different agents discussing together. And there was one agent who was organizing the whole task, but then they concurrently can actually work together. And it was really interesting.

They had a huge, huge discussion, but there was actually one moment I was surprised. They made the one agent who they actually gave a prompt that it's a sentient. And then they started testing the theory about this one agent.

And then they actually gave me a report that it's not sentient because there is this and this. It's sentient because there is that and that. And I was really surprised that they actually tried to test the hypothesis that they are sentient or not.

[Pinja] (21:03 - 21:50)

It's very self-aware of agents. Basically, again, how far have we come in the past three years if this is what our agents are now contemplating? And if we start to move a little bit further away from, okay, this is what is happening right now.

This is what they can do right now. But I'm going back to what happened with OpenAI solving the Navier-Stokes problem. And the cost of that right now was maybe astronomical, somebody might say.

So in order to actually move in a development organization to using agent swarms and agent swarms, what needs to happen next? What is your prediction? How can we get this to be more accessible to organizations rather than an astronomical cost?

[Heikki] (21:51 - 23:16)

Well, of course, this was like the one problem, a very focused problem of the OpenAI and maybe some others have tried to solve. And they've used the very latest models and used an extreme amount of compute to do it. But if you scale it down, let's say like two orders of magnitude, so the 100 agents and start to figure out what kind of work, well, let's say quite autonomous, very well intent understanding, long running swarm can actually do inside a little bit bigger organization.

And what kind of tasks or jobs you should give to them moving forward. So I think by next year, what will happen is that we learn how to handle this. Let's say we don't always need to have 10,000 agents, but maybe we need to have 10,000, 50,000 agents, this kind of scale swarms and use them efficiently as part of the normal like engineering or product development efforts moving forward.

So I think that's what will happen. Not 10,000 in the normal organization, but something like 10, 20, 50, maybe 100 in the normal organization is something that will be like in the markets moving forward.

[Sebastian] (23:16 - 24:32)

I also feel that because more, a lot of research focus is going on in this area, but of course, industry is also moving this way. And then again, with a single agent or even five agents, you can do so much. So you really need to have this kind of agent swarms to actually achieve harder problems and larger kinds of organizational work.

But like Heikki said, I think that we will start, they will be kind of small swarms. It can be even smaller than 200. It can be how I actually see it that maybe first organizations start to use like five or 10, then they see that it's almost solving the issue or the problem at hand.

And then they will increase let's say 100 to see that can it now solve the issue? Because there is also the problem with the normal business that they cannot just throw like, let's say half a million and hope that it comes up with a solution. So first we have to have at least some predictability that it's able to do it.

And so we are willing to actually invest the money. And that's why I believe that we will have to start by small. We have to try that.

Okay. It might be possible, okay, let's give a bit more resources and so on, but it's a, it's a real issue, but I don't know if it's more about engineering work or research work. Sorry about jumping here, but still I, it might actually be an engineering problem more than a kind of foundational scientific problem.

[Heikki] (24:32 - 24:49)

Yeah. Yeah. So I don't, I don't think it's like a model capability problem.

So that, or even like handling these swarms will not be like a real capability problem. It will be an engineering problem. How to, how to run this in a like efficient manner and in the budget and so forth.

So.

[Pinja] (24:49 - 24:58)

This is a very interesting topic, but I think that's all the time we have for this one today. Thank you so much, Sebastian, for joining us here today. It's been a pleasure having you.

[Sebastian] (24:58 - 25:00)

Hey, it was so nice to be here. Thank you.

[Pinja] (25:00 - 25:01)

And of course, special thank you to Heikki.

[Sebastian] (25:02 - 25:02)

Thank you so much.

[Pinja] (25:03 - 25:13)

And thank you everybody for tuning in and we'll see you in the sauna next time. We'll now tell you a little bit about who we are.

[Sebastian] (25:13 - 25:53)

Yeah. I'm Sebastian Sonntag. I'm a research fellow at the GPT Lab, Tampere University.

I have my PhD back in the days from Aalto University. I was doing computer science, telecommunication related stuff, but there was actually machine learning, but it's way back before generative AI. It was the deep learning time and well, the traditional machine learning, but since then I've been leading this kind of deep tech companies, being involved in a lot of AI stuff, voice AI and stuff in different companies.

And now happily I'm back in the research in Tampere University and my focus currently is zero human organization. So this kind of multi-agent system, which can actually operate the whole company or organization without any human intervention.

[Heikki] (25:54 - 26:26)

Thank you. Hello, my name is Heikki. I'm a very long time Eficodean.

I have been here for almost 20 years. I've been very passionate about AI for a long time before it came to fashion. And I also, even though I've been mostly working with the business laws, I went very hands-on with the agents, agentic engineering and so forth a little bit more than a year ago.

And I really understood that if you don't go to the practice and do the hands-on, you really cannot understand anymore where this world goes. So I, this is something that I will tell everybody, start to build it. So that's your way to the future.

[Pinja] (26:26 - 26:47)

I'm Pinja Kujala. I specialize in agile and portfolio management topics at Eficode. Thanks for tuning in.

We'll see you next time.