Reverse Conway's law and GenAI: How agents will take over the organisation - Patrick Debois

Today we're going to talk about Conway's law, Reverse Conway's law, and generative AI, and it's basically thinking about the future of how AI will influence our organizations. Now, we got this new technology,right, like ChatGPT was the first, and it, like, impacted—like everywhere—the whole industry.

And this really brought, like, everybody was, like, so excited that, like, they kind of changed everything to becoming more of an AI specialist. And that's really cool because everybody's still learning, and we'll see where we end up. Now, some kind of just sprinkle some AI on there, and they don't go all the way, and that's fine, but I don't think they will kind of really get the results that they want from it.

Now, the bigger questions that companies who do want GenAI, they wanted to know, like, where does it land in the organization? Will it belong to the data team, the application team, the platform team? We don't really know. So we already see that, like, shaping our org for this new technology is something that is happening in a lot of organizations.

Now, last year at the AI Engineer event, I talked about platform teams and how kind of you can introduce, kind of, and scale AI out of. You can watch the video here, but that's not what we're going to talk about today.

Today we're going to talk about Conway's law, and simplistically it says that whatever way you organize yourself and your teams will have an impact on what you're able to build. Like, if you have small distributed teams, you're likely going to go with modular services and architecture.

Here you can see, like, various pieces of organization, and a lot has to do with communication lines and how the communication lines actually influence parts of the organization. So there's a kind of impact of what we can build.

Now, who am I to talk about this? Well, I've been in automation for a long time, been following a lot of the DevOps, and now being really focused on doing more with GenAI, as anybody else else,right. And I'm particularly focused on engineering more with AI, so kind of as a software developer to do more with AI and kind of bring the benefits out.

If you're looking at this video, if you're interested in the slides afterwards, please hit me a message on LinkedIn. I'll be happy to share in exchange for some feedback or things to improve. Another thing that I'm active on is the AI Native Dev Community.

I'm a curator and a contributor where we try to get the news stories about how kind of development will change inside of the organizations. Now, I told you a little bit of how we're organizing for AI with a platform team, but today we are doing the opposite.

How is AI changing the way we organize ourselves? So that's kind of a kind of a kicker or a kickback in this. Now, what have you seen so far? One of the obvious things is, first level is everybody kind of starts accepting AI as a copilot, and that's really great.

Some kind of start going even further. They think about, like, training or their knowledge. This is came back, and, like, all the knowledge that we have, it is changing, and we can pass that on to that sidekick. That's great.

Now, eventually, we're already seeing that AI is writing more and more code and will become more of a contributor. So even though it's a sidekick and kind of a copilot, it really drives us to more efficiency and more contributions in the organization.

Now, this brings the elephant in the room. Like, are we still needed? And that's the question a lot of people ask. What should we do? Will they eventually kind of change the way we have to operate? Now, that's normal, and a lot of people are kind of, like, dealing with this issue at work, and they are scared.

And often you have to think, like, when a new technology comes, you have to rethink your job, and sometimes the culture of your company really supports that, and you're able to kind of help and learn about that news.

I hope that's the case for you and that you actually bring this new skill of how to figure out what you need to do in the next couple of years. Now, what we're seeing is, instead of just coding, as an example, we're moving away from just typing and having it help generate some of the text to a new role is that we start thinking about the intent, something we've always wanted, but we've been there kind of in the toil, writing code, typing, and now kind of the new technology is taking this up to a level where we can think more about the planning and the intent.

There's even products that start emerging, like product requirements, and that generates the code, so we're less about the typing anymore. Or you could say it's becoming specification-centric, and we feed it with the specs, and then eventually it builds the system.

Now, it's almost like, well, we're reviewing, and we still have to be good at reviewing. We say, this is good, this is bad, but we're going to do less typing. And this is just an example of what everybody kind of using copilots and so on is kind of experiencing.

You become from, you know, kind of more the creator to the person that is managing and kind of more on the how, sorry, more on the why and less on the how. And that's great. Now, to put this in perspective, every time there's a new technology, like, people already have a job, and they do a bundle of tasks, as the upper kind of shows.

Now, technology comes along, sometimes it adds, or it's the copilot, or it kind of complements what we do, and sometimes it will be substituted. So kind of that unbundling of tasks is something that is normal when you kind of, like, move with a new technology.

Now, what can happen? It's just going to be the same. We're just going to do more, and we ask kind of extra for this,right? So sometimes that's what technology does. It adds a premium to what we can do.

Now, sometimes, like, maybe the value that we put in coding is moved somewhere else to management or kind of manage specifications, and that's kind of another thing that might be happening. So even though you are, like, moving away from your existing tasks, something happens in another tasks.

And then eventually it might be commodity, and that's where you kind of, like, get this feeling of, hey, if this is commodity, what should I really be doing, and how to go into another task to kind of prove your value to the system.

And eventually that piece that was kind of really good and kind of was delegated becomes substituted. We don't care anymore. The value is kind of not good anymore, and we really have to find, like, another job in the system.

Those things are the typical scenarios a new technology brings. Now, you might think, like, maybe we're not coding anymore, but we're training the AI to do coding. Here's an example outlier where, you know, people get paid. I'm not saying this is a bad job, but it shows you kind of how technology can change some of the tasks and the different kind of jobs that are being created.

Now, eventually you will see kind of these stages. It won't affect me. That's when kind of the task is kind of complementary. And then you kind of go all the way up to, hey, it's going to replace me.

That's really how. And eventually you find a spot that says, I can now do things that I couldn't do before, and I have now new superpowers, and I can create something. So kind of that's the roller coaster that we're on as an individual person in this new technology.

So this is how AI is affecting us as an individual. Now, we know that things will go wrong. So in this case, Amazon too bought kind of, like, you know, increasing the price because one is bidding over the other.

So humans are still needed,right, in this case. But as I showed before, it might not be in all cases, but things will go wrong, and we have to be prepared for that. And sometimes humans just have to clean up because that's what we do.

We understand what went wrong. We make sure that, like, the machine can't handle it. And this is the intermediate scenario that we're going for. Now, a lot of the automation, what I learned is that a lot of it is about automation, and then it is about preventing failure and designing from failure and so on.

So kind of dealing with failure is still typically a human job. It might require less people, but when you kind of have this happening, you have to be prepared. So you still have to learn, train, and deal with all kind of the situations.

So it's almost like a paradox. Like, the more we gained with the automation, we still have to be training and be ready when the failure happens. So it could be kind of a net zero effect as well. Now, let's move up.

We kind of have that delegated assistant copilot, and now we're heading towards the team level. And instead of having a personal assistant, we think about, like, a team member that has access to the team that helps the whole team to do a better job.

And sometimes the impact at the team level is the domain knowledge. And as you can see, when multiple people are on a team, they have various levels of domains, and we really focused on having kind of multi-purpose or multi-domain knowledgeable people, the full stack engineer, and bringing a team of experts of various domains together.

We found out that's the best way to organize the team, and that's great. Now, if you think about, like, the human team, that's great. It has a limited number of knowledge, which is contained, let's say, the seven or eight team members, and then there's the whole domain knowledge that an AI can have.

It is way bigger, and it can now be another way of getting that knowledge. You already see that, like, people getting questions, asking it, but it's still the human that does the verification, but it's vastly helpful to this more automated as an agent.

So eventually this might lead to smaller teams because that overlapping is maybe not that needed because we have a system that actually knows a lot. Again, we have to still understand what good looks like. So that might change the skills that we require in those teams as well.

And kind of maybe those smaller teams again become another team. Who knows? Kind of like that is a scaling issue. We're kind of all speculated. Some I've seen some cases where teams got smaller. I've seen some cases where teams got bigger.

We don't really know, but the dynamic will likely change with more AI being as a team member. And you can say that the stack that we're building, like, whatever we want to know, there's the applications, there's more of the infrastructure, and then the teammates, and we're looking for more domain-specific teammates, much like the diversity that we brought in teams that can help somebody in sales.

Somebody is really good at data. Somebody is really good at other skills. So kind of bringing a new teammate stack altogether with the technology stack. So kind of that expansion is something that you'll see happen with more, you know, agents and parts becoming part of the team.

Now, that means that, as I mentioned before, that we might be faster, and they can do things a lot faster, but you are still the person who kind of has to make the decision. So maybe there's not enough work for you, but there's a lot that you can do with multiple of these assistants and team members.

And that's kind of a specialist that goes in when things fail or has an understanding of what to do. Now, what we all see is that maybe the cycles that were used to be the one or two weeks get reused in two days.

So it's becoming a shorter feedback cycle. So that means we're getting into the loop, but we can't have, like, the conversations with everybody. So there's going to be multiple feedback cycles, maybe that complete in a bigger feedback cycle, but we're getting to a more real-time cycle compared to the typical kind of one or two weeks or three weeks that was required by humans to do something.

Now, we're all going to learn. And as I mentioned, the new way of forming a team is maybe more of kind of the team members that are agents, and that leaves that one person not just to be the specialist, but they might have to be switching between domains because they understand A, and then tomorrow they move to another project, and they kind of work on that.

And so kind of that flexibility or fluid way of working is something you do. So this doesn't come overnight, and people have to train for this. But again, that might be the way that we're going because they are becoming more of the overseers in the system.

And so AI is there to support us. And when AI gets more, we trust AI more, or they become more trustworthy. Maybe kind of they take away some of the jobs that we already have as a team member, and it goes away.

So that kind of shift from the pyramid above to kind of like the pyramid below, that kind of the human section moves up and might become smaller as well. Now, we can already see this in another way helping that, like, the system is helping new developers join a team.

And that's great because the learning curve has been reduced. We need less interactions. It's more tailored. It's more personalized. That's one way of helping new people in the team. And one might argue that the point of this presentation, do we still have the true understanding?

But as I mentioned before, if you don't understand, how can you make decisions? How can you make the call? How can you deal with failure if you don't truly understand? Yes, the systems might help you. They might have a large expertise, but we know that these systems inherently have failures.

So we have to be prepared as well. So I think we still need to understand the thing, maybe not on a daily level, but if we're the person that has to deal with the issues, we still have to understand.

And kind of making decisions means you still understand what goods look like.

And then people often argue about, like, well, the juniors might not get into the field anymore because only the seniors know what goods look like. But we can have an accelerated learning there as well that a junior becomes a senior faster thanks to that AI and thanks to the training and the more personal things.

So maybe that doesn't have really an impact on the juniors, but I'm sure it might have an effect kind of psychology-wise that you might not be entering the field because it feels daunting to be, you know, instant having to be the specialist as well.

So we'll have more training programs that will help that as well. And you already see that kind of shrinking of a company the more we delegate. Maybe before it was more delegating to SOS companies, but now we're delegating to assistants, and the companies might become smaller, much like the teams become smaller, and we work from there to kind of have a different pattern on how we organize ourselves in a team.

Allright. So in this case, in the previous case, we were still about, like, having an assistant and it helping us was part of the team. It took away some of the jobs, but now we're entering the level. Well, if it can do certain tasks as good as a human, it might become a peer.

And so that level of rising from being a teammate to a peer on kind of helping us out. Now, why is that not very strange? Because these LLMs, they contain a lot of culture of how we work together.

And that's kind of how these agents will mimic. Kind of that is the way that culture is transmitted, much like Wikipedia transmitted a lot of knowledge. These kind of models can help us, kind of help us bring a lot of knowledge into the system as well.

And it's not crazy. This might be the first paper about it where they kind of have multiple agents walking the town, and they saw all the behaviors of one team talking to the other team. Guess what? They also learned that, you know, when they collaborated, when they talked more, they get better results.

So it definitely mimics part of the human behavior in this as well. And then one thing we hope for if they become a peer is that they can go off and do things on our behalf or like as a peer that they can help us.

So kind of that is probably where a lot of people hope it's going, but it's going to be challenging because when do we get that trust level as well? And why not, like, have a whole company on this?

You know, you have a software company, an investor, a product manager, and so on. So you can simulate the whole process with agents, and that's really cool. Would I trust itright now? No, but it's definitely interesting to see where this ends up.

Now, we mentioned about kind of having smaller teams, but this also has an impact on the organization. So maybe there's going to be less leadership, and it's just going to use a bunch of AI agents, and the teams get smaller indeed.

Now, agents might have unwanted behavior, much like humans, is that we really want to control what happens if they do certain things. Some have started talking about toxic behavior of an agent, like one agent is influencing other agents to do bad things.

Does it remind, doesn't it remind you of humans, like one bad apple impacting other team members? It's exactly the similar behavior. So we'll have to watch out for this kind of agent dynamics, much like we did with team dynamics as well.

And, you know, is AI still going to be inclusive for humans? We'll have to watch out. Like in this example, you walk into a supermarket, and if you don't smile, you don't get in. And the AI is deciding what happens to you and whether you're allowed to go in.

And kind of that level is interesting to see how kind of AI will keep us inclusive.

OpenAI has also been working on kind of what they call a model spec. So think of it. What an LLM is allowed to do, it's the guardrails, it's the rules that you're allowed to do. And it's very similar to a code of conduct for humans working in a team or in an organization.

So we're mimicking the same kind of behavior for those agents as well. And we want to be inclusive as well, even though it's technology, people fear like, hey, it's going to get feelings. I don't think so, but it doesn't mean that we have to be really bad to the systems because that might influence some of the toxic behavior.

And it's still, if humans can influence what the agents do, it's the same as bad agents influencing kind of what the whole team of agents do. We're going to maybe see how the market evolves from more general agents to specialized agents and handcrafted agents.

Maybe one is really good at a certain task. As we mentioned before, one is not available and is very niche. So we're going to get more handcrafted agents, and some might work all the time, some are very specialists, and it is very similar to kind of the specialists versus the generalists of a human working in a team that we kind of have a different rewarding system as well and how it will be trained on the tasks.

Allright. This is maybe a stupid thing they did, but it was almost like every agent could get an employee record. Like, think of it as equal rights. This was redrawn. There's a lot of backfire on this. Obviously, that's not what people want.

It's not human. It's not there. But the question is, how do we deal with this as well? And if we take this up a notch, could this be a manager of people? I showed you like who gets into the store.

That's a type of manager. That's a type of decision that we're delegating to the agents already. Now, that brings up a question like, where do we allow an agent to sit in the organization? A peer, like an assistant, a peer, and we go up all the way up to a manager.

So that again brings that AI inclusiveness. Like, what is the decisions? And we have to have kind of all the different code of conducts, guardrails, and everything in place and still have to understand what the failure is. So we're still having a human having to manage this as well.

Then you can think about like, hey, if some of those pieces get done by the agents, we can put them next into the org chart. Interesting way of thinking about this. Like, what is the equivalent of one agent?

Is it like five humans? What's the cost? So it brings a little bit of the return of investment when you're planning an organization. And sometimes that's not really clear, but we're going to see more of those kind of discussions like, well, should I hire you because I have a team of agents that has the same output?

So that kind of discussions are about to be happening already. And then maybe, you know, it's the agents that go to our job here. They kind of like, you know, applying for LinkedIn jobs. Maybe we don't work, but we have managing a team of agents that actually do the work for us and help.

So we're becoming more of a task force on our own in that perspective.

And yeah, we're working. The AIs are working for us and that, you know, might have a different LinkedIn of people finding work and finding people and services that they want to use. Now, if they're working as us and they're in the HR and we treat them as humans kind of and the performance, then we'll have to do performance reviews as well.

Like, are they bringing the value or kind of should we kind of give them different incentives? Much like humans get incentives, there's the individual incentive, the team incentive, and the business incentive. How do we kind of steer that clear with the agents as well?

So kind of that ROI driven with agents is going to be way more specific than probably for humans, where we don't want to be that like that clear on, hey, you're doing this, this is the return. Some domains really do this, but we're going to see way more of this when agents do certain parts of our job or take over like teams as well.

And if you think about this, we can like, much like we do simulations when somebody joins a team as a human, we can run this like, what's the impact of a new person hiring as a new agent joining our team?

So kind of like that kind of experiments is what people are trying to figure out with a digital twin of the organization already. Now, we might need an API. That's just semantics. We understand what the org chart is, HR, and it's all very sensitive often.

So it might not happen, but it's definitely like a thought experiment to kind of think again on the ROI of hiring people, hiring agents in that perspective already. And then the question becomes like, who controls who? Is it us controlling the agents, the agents controlling us?

We're not sure. And kind of that's, you know, a very weird situation to be in because ultimately who watches the watchers is where we're heading to already. And then eventually maybe they become the business,right? Which is really strange in this.

Like if you're building software and you have a team that builds software and that's all good and you sell that software, now all of a sudden this becomes

different if, let's say, we're used to kind of buy services from a third-party company to kind of have very good service. If everything becomes so cheap, I might do internal labor again through my agents and the agents building systems for me.

So it might be that disruptive that we're not doing software as a service, but kind of like bringing kind of service as a software in there already. And so we see how we went through it, like the individual impact level, the team, and then we kind of start replacing us and maybe kind of we do the work and they start doing the work and we're going for.

So kind of that is the speculation. There's no kind of, you know, clear-cut path. But I thought in this presentation, I show you different forms how people are thinking about that stuff and how we're going to get to that level.

Now, hyperhumans, who knows, kind of that combination of a human starting to sound like a little bit cyborg. I know I have no clue, but something is happening. Something is changing in our organizations. And it will all depend on how good the technology will become and how much trust we can put in there already.

And yeah, I agree. Fascinating stuff. I don't know, but I agree. It is something we have to think about. It sounds a little bit like science fiction, but it might be there like faster than we think, and the impact might be something we don't see coming that good.

And while many say, you know, AI won't take my job, but somebody using AI will take my job. Well, maybe there's good parts that will take away the toil and so on. But obviously the whole discussion becomes like, how will we get paid from there?

And like, how will our economy thrive if we go towards more of those agents? And my advice is you have to think about like, stop building the thing, like don't build the software yourself, but kind of think about like how you would build the thing, how you will kind of build the AI that builds the things that you're currently building.

And I think that's the best way to kind of go up a middle level and deal with this, because then you actually understand what the things are doing. You are very good positioned to do supervision, and we also kind of bring the value to the system with a new skill to do so.

Thank you very much. And I hope you enjoyed this more futuristic talk about like the impact of AI on the organization and kind of on us as an individual. You can watch most of the similar videos on my YouTube channel or kind of connect on LinkedIn and please leave feedback and let me know and enjoy the rest of the videos.

Thank you very much.