What are Akgents? 

In short, Akgents, or actor-agents, are an operational implementation of the AI Transformation value proposition, about which you can read more here. 

Akgents are a new twist on an old idea. Back in the ‘70s people were thinking about programming paradigms suitable for parallelising computational tasks, and along came the idea of the Actor Model. If you’ve run across the Actor Model in the past couple of decades it was probably not in the context of AI, but the focus of the original paper was in fact about artificial intelligence which is even mentioned in the title:  “A Universal Modular ACTOR Formalism for Artificial Intelligence.” ¹ I find it fascinating that as far back as the 1970s people were thinking about how to build robust, scalable AI systems. If you read the paper, which I highly recommend, you can see that they went quite far (at least in theory)! 

Over time, the idea of actors took back stage to concepts such as object oriented programming, functional programming etc, but I believe it’s about to make a huge comeback. 

The reason, you’re probably guessing, is multi-agent AI systems!

The beauty of the Actor Model is in the simplicity! Actors are logical entities which can do four things: 

• Maintain their own state
• Perform actions
• Send messages to other actors
• Spawn/destroy other actors

That’s it. That’s the whole model. As it turns out, this is enough to build robust AI systems which solve many of the issues AI agents are currently facing. 

‍The approach does not require any intrinsic structure or process, but supports these concepts as a part of design or even as an emergent property! See for instance the debate example below: it is pure functional chaos, but it works! 

At B12 and Yuma, we have been thinking hard about how to go beyond the current SOTA in agentic AI systems. We believe that multi-agent systems, which allow us to form teams of AI agents and humans in a seamless, safe and efficient manner will be the next big thing in AI. However, frameworks such as Langchain, while great for building back office applications and AI powered workflows suffer from certain technical limitations in the context of multi-agents. For instance, Langchain graphs are static, meaning that if you want to add or take out agents, you have to generate a new graph, reload and restart. Hardly desirable if you are building a scalable application. In addition, if you want to put the human in the loop, you have to stop the execution of the entire graph. Given that humans and AI work on vastly different timescales, this is not optimal. Finally, nodes in a lang-chain share a global state which is cumbersome, from many aspects: governance, scalability, etc.

This is why we started thinking about other approaches, and the actor model seemed like a natural avenue to explore. A few intense whiteboard discussions later, the whole thing started to quickly click into place, resulting in a new approach based on the actor model. Recently, we also learned that Microsoft is working on a similar approach to multi-agents, with the AutoGen 0.4 version being based on the actor model as well. We saw this as an exciting validation of our ideas!

We were inspired by the idea of actors to elevate the notion of an agent to a logical entity which “lives” inside an actor. This can be a Langraph agent with multiple tools for instance, a deterministic piece of code, or a human being! Each actor can be powered by its own AI model, have a defined role, personality, tools, actions and rules about who it can send messages to and when. Agents can live in real space (e.g. us humans), in the cloud, or on premise. There is no limitation to where you can deploy them!

‍The beauty of the framework is that it is less about technology and more about the organisation aspects of the Akgent team, which emerges as the central concept. This is very important because when we talk about multi-agent systems, AI, software etc. we don’t really care about the details of technology. We care about how we can use this technology to do something useful with it. The Akgent framework allows you to shift your mindset from technical details to the core objectives in a very organic way.

Akgents directly address the five pillars of AI transformation described above:

1. Holistic Perspective: Akgents act as team members in any situation and can be adapted to any organisation or process, and can be used to redefine the core notions of an organisation at any level.
   
2. Human Empowerment: Akgents are designed to elevate the role of the human, not replace them! The approach is intrinsically inclusive, making change management easier, as well as resulting in superior value generation compared to, say, the full automation approach to AI.
   
3. Collaborative intelligence: The core concept of the approach is not the agent itself but the team! Akgents solve a prominent technical problem of allowing AI agents and human agents to seamlessly communicate without needing to interrupt operations, despite the fact that humans and AI work on different timescales, with the ultimate goal of allowing AI and humans to function as more than the mere sum of the parts.‍
4. Sovereignty: Akgents can be powered by any AI model, anywhere! This means that management of governance issues is built into the framework. If a part of the multi-agent team needs to live in a protected data space, this can easily be accommodated. If another part needs to be in the cloud, so be it.
5. Sustainability & Resilience: Akgentic teams are architecturally elastic (if one akgent fails, the system keeps going). They also auto-scale well - you can create and destroy Akgents on the fly, utilising resources when you need to and ensuring a high degree of scalability and green computing practices. 

Let’s see a few examples of what you can do with the Akgents.

A Presidential Debate with a Puppeteer Agent

Say we want to simulate a realistic presidential debate by using AI powered agents. You need a moderator who will generate questions on a topic and then ensure that both candidates have been given a chance to answer them. The two candidates will answer the moderator’s questions as well as react to their opponent’s answers. 

So far, the structure is something you can do with pure Python and access to an OpenAI API. But let’s say that now we want to be able to add an audience into the mix. Now it gets interesting.

The audience is there to listen, but can also react to what they are hearing. If they support one candidate they can shout in support, but if they hear the other candidate they can boo him/her. On top of that, we don’t really know in real life when the audience will react or how exactly. The debaters also hear the audience and react to them. This is something that is very difficult to do without an inherently asynchronous multi-agent system. 

Let’s go a step further. I, a devious human agent, can control who gets an audience, how many of them and what their agenda should be. This has to happen without affecting the flow of the debate. In other words, the audience agents should spawn (or disappear), start listening, and respond to what they hear without the debaters, moderator or other audience members flinching. 

I can also communicate with the moderator agent by “whispering” instructions to it on the fly, without affecting the flow of the debate. For instance, I can instruct it to change the topic of the debate to whatever I desire. Finally, I can modify the personality (i.e. state) of any agent while the debate is running. 

Sounds crazy, right? But it is possible!

Watch the video below to see how it works.

‍The debate example serves to illustrate how you can build a useful, functioning multi-agent system without any need for a pre-defined process or structure!

You could simulate any team structure, or how to design information flow through an organisational structure in an optimal way. You can do this while taking into account aspects such as acting on partial information, adding/subtracting team members, changing personalities, objectives and tasks in the middle of the process etc. It’s AI powered management consulting! 

Support Team

A very common challenge many organisations face is managing data processing. Whether you are a law firm, an insurance company or a manufacturer of consumer goods, you very likely have to deal with a large amount of emails, documents etc. on a daily basis, and processing of this data can add up to a very resource intensive operation.

To make things even more complex, it is commonly difficult to predict what the content or purpose of incoming emails or documents will be. A client may ask you about your product, but also about a different issue. They can also provide one or more documents of varying content and format. Attempting to solve this challenge by building data processing pipelines is doomed to fail, as the overall  complexity of scenarios you have to handle is simply too large to allow for a pipeline-like system which is scalable or maintainable. 

Akgents offer a way out! Instead of focusing on rigid, process driven data intake, Akgents allow you to shift your mindset from the task of data processing, to the team objective of resolving a customer issue, providing a report, etc. Imagine for example that you run a customer support center for a SaaS company and that you receive emails from your customers on a daily basis. These emails can be about any feature of your SaaS, pricing, questions about your development roadmap, regulatory inquiries etc. They can also be simple end of the year congratulations. 

Fully automating this process can, of course, be risky, so you should keep the human in the loop. You can imagine a hybrid human-AI system, where an Email Analyst Agent receives the incoming email and determines which questions need to be addressed (and perhaps which documents are needed to address it). For each question, the Email Analyst can delegate the process of formulating the answer, by spawning Support Team Members. Support Team Members can have access to knowledge bases via RAG systems or other ways, and can determine if there is missing information. If information is missing they can relay the message back to the Analyst, who can either decide to email the client back asking for the missing info, or refer to the human in the team. The Support Team Members can wait for the extra information to arrive, but if another email comes, the Senior Analyst can simply spawn another team of Juniors to process it without affecting the state of affairs for the other email. The human on the other hand can choose to intervene and provide new information to any analyst, or forward the customer information. They can also choose to change the behavior of the analyst on the fly if they notice that, say, an analyst is repeating a mistake due to  lack of clarity in the context.

During execution, agents can ask the human questions, and the human can interfere with the information flow by chatting with the agents, in the same ways as you would send instructions to a junior team member if they did something wrong. 

Here is a video of a multi-agent email  and you can see the video demonstration of how it works here:

The whole system very much resembles how human teams work to begin with, and it is precisely this intuitive approach to AI system design that is a huge part of the value of the Akgent approach - at the end of the day the problem you have to solve is less technological and more organisational. 

‍Why not handle this with a single Email Analyst agent who handles everything? You could, but you would end up with a system which is a lot more resource demanding, more difficult to design from the governance perspective and overall less resilient.

AI Powered (and not only video) Games

Video games are a very natural environment for multi-agent systems, which has potential to make games a lot more realistic, non-linear and more interesting. Imagine that you are building a world exploration role playing game where the human-driven characters are encountering other, AI powered characters. You can imagine that each of them is an Akgent, with their own state, ability to use tools, perform actions and communicate. Each type of Akgent can have a personality. This personality can evolve over time, either through human input, or through the changes in the agent’s state (in real life we would call this experience). It would be very difficult to predict the course of such a game, but this may be what will make games more fun in the future! 

Akgent powered games can serve as more realistic emulations of social systems beyond what was possible in the past. We can use them not only to build more interesting video games, but to simulate organisational structures or devise AI driven strategies for resilient supply chains. 

Conclusions

To summarise, Akgents offer a new paradigm to build AI powered multi-agent systems which don’t suffer from many of the drawbacks of current frameworks such as a global state and lack of ability for concurrent execution. They allow for seamless integration of AI agents and humans, in ways which respect the fact that each of them operates at different time scales. Finally, the fact that we can create/destroy Akgents when we need to allow for constructing much more scalable AI systems than before. We are very excited about the possibilities this framework will bring to build new generation multi-agent systems! 

‍

References

1. https://www.ijcai.org/Proceedings/73/Papers/027B.pdf

‍

About the author

‍Mihailo Backović started as a theoretical physicist turned AI expert who specializes in guiding organizations through successful AI transformations. As a Managing Partner at B12 Consulting | part of Yuma, he is helping numerous companies—from biotech and healthcare to mobility and finance—harness AI to achieve their biggest ambitions. Known for his deep analytical approach and passion for tackling complex problems, Mihailo is dedicated to shaping a future where AI and humanity thrive together. He actively explores the transformative impact of AI on the topics such as the job market, democracy, disinformation, and is committed to building AI solutions that empower people and businesses to flourish in an increasingly automated world.