Many technologies are advertised as “agentic” right now. But what actually makes a technology agentic, and is it correct to call a simple chatbot already agentic? To answer these questions and talk about his latest project, we sat down with our Data Scientist Neil Sinclair in our Talk Data To Me interview series.
You would like to watch this interview instead of reading it? Then go here.
Antonia Mittmann: Neil, can you please introduce yourself?
Neil Sinclair: My name is Neil and I’m a Senior Data Scientist at the Berlin Data Studio. I’ve been here for three years. In the last two years, I’ve worked on a lot of really cool genAI projects, including the one I’m going to talk about later.
Source: applydata.
What Is an Agentic AI System?
Okay, Neil, to start: what makes an AI an agent, and why
is a system called agentic? How would you explain that?
I think one would want to make a distinction between a normal chatbot and agentic AI. I would argue that a chatbot or any language model alone is an agent. And agents can be more or less complex. You can give some tools to an agent, for example, access to a database or maybe a tool to allow it to write and run some code.
But an agentic system has multiple agents operating together. What makes these systems so great is that you’re going to have agents overseeing the work of others and giving feedback to each other. And you can also let these agents have tools so they can start to orchestrate workflows. In this process of them communicating with each other, we start to see real intelligence emerging from these systems.
How Does an Agentic AI System Work?
That sounds super cool but also super complex. Could you give
me a specific example of these agents working together?
Sure. A good example of this might be a meal-planning app or system. Imagine the weekend’s coming up, it’s autumn, and you want something to help you plan three dinners and two breakfasts. You talk to a chatbot and say, “Hey, the weekend’s coming up, I’m a vegetarian, and it’s autumn. Can you help me plan three breakfasts and two dinners?” This agent would then take your information and pass it to another agent that would break it down into some search terms: “autumn”, “autumnal foods”, “vegetarian dinners”, “breakfasts”, things like this. That agent might then pass it to another agent that can search in the database and get that information back.
We could then, along with your original query and that information, pass it to another agent that could say, “Does this information from the query match what the user wants?” If it doesn’t, it could send a message back to the search agent and ask it to search again until it finds better information.
Maybe it passes that information back to you and says, “How do these meals look?” You say, “Great.” This thing gets passed to another agent that maybe has an API connection to some sort of food-ordering service. It could then break down your meals into ingredients, send them through an API to the food ordering service, and get the ingredients back.
Maybe it will want to check with another agent: “Have we managed to order everything for these meals that we want? Or were there any ingredients we couldn’t get?” This checking agent could then say, “Okay, I see that they don’t have this kind of cheese”. You could get a different kind of cheese, and it sends this information back to that agent. It can make the order and then send you the final order to approve.
So we have all of this intelligence going on in the background, with these agents speaking to each other. You basically just have to tick off one, two things: “Yeah, I like these meals”, “Yeah, I’m happy with this order”. Like some workers who communicate, and there’s not one agent that can do it all, but they’re somehow intelligently splitting the work.
Now that we’ve cleared up how it works, I’d like to know a bit more about your
latest project. It’s called NEVO, and it’s a digital sales assistant that also
uses the agentic AI you just explained. Please introduce me to this project.
The idea with NEVO was that we would make something intelligent, personalized, and scalable. NEVO is a voice-activated sales agent that takes a user through the sales journey and generates specific content for them based on what they’re speaking about and what they’re interested in.
Which AI Agents Exist?
That’s super interesting because I’d say it’s a time where personalization is something every
customer demands, especially in online shops. I’m not too much into code
and not too technical, but maybe you can explain a little bit how NEVO works?
Which agentic assistants do we have in this system?
We’ve got one core agent that’s listening to what the user is saying and then passing the user to different experts along the customer journey. So, different experts have access to different bits of information. There are other agents in the background that are also listening to the conversation and extracting key information that they could then pass to a human salesperson later.
And some of the really interesting stuff is: we have these agents that are selecting specific content – in our case, specific photographs. Then we have agents that are writing specific content for the user based on some general information and are then conditioned on what information the user has been sharing with them. This makes these really personalized descriptions of the products based on what the users are interested in.
…Sounds like I would not have to read four pages of unnecessary information
about a product when I just want to know if, for example, the coffee machine
fits in my kitchen, and I don’t want to see every detail of the machine
when I’m not an expert at all.
What’s Important when Building Agentic AI?
I understand now the back end, but we also have the human: the customer.
The human communicates with this agentic system.
What was important for you when designing this communication?
I think it depends on who you ask about the project. Different people thought different things were the most important. Something I thought was really important was that as you go through the customer journey, you’re able to go forward but also backward. So when you buy something in a store, it isn’t like you always just walk in and pick up the item and pay for it and walk out, right? Sometimes you’re going to look at something and decide between different things. Maybe something goes in the basket, goes out of the basket. And we needed to create this intelligence in the system so that a user would be able to go forward and backward in the sales journey as they learn new information about the products that they’re interested in.
When you were working on the project, which part of the
whole development process did you enjoy the most?
There are these limitations or constraints that we have with using a service like Gemini or OpenAI. And one of the really cool things that we did inside of these constraints was set up some things that we had NEVO reacting in almost real time. This was very technical, but it was so much fun to get right.
Challenges in GenAI Projects
Speaking of challenges, was there another technical challenge or maybe also
a project challenge, anything you encountered that you had to deal with?
Yeah, with this project and every genAI project that I’ve worked on, one of the biggest challenges is assessing the system because a small change to the input can lead to a big change in the output. And especially when you have multiple agents working together, a really small change in the input can lead to something completely different in the output. It’s not too difficult to test the different agents or the individual agents. But to test all of them together can get quite complex.
One of the solutions that we came up with was to create an AI voice-first customer that would speak to NEVO, and we would watch these interactions between them – either listen to the interaction or read the text of the interaction afterwards – to decide how to change the system. That was quite a lot of fun.
It also sounds like fun! And it uses all the potential of AI, if you can even
useit for testing the AI. To sum it up, which limitations of agentic AI
ingeneral do you still see, and how could you address them?
I think one of the challenges is this brittleness of the system: a small change in the input can lead to a big change in the output. I think humans are quite good at being adaptable when things are a little bit different when they come in. Language models are not so good at that yet. However, as we create these systems that are able to give more feedback inside the system, or maybe just get humans involved in specific parts when the system’s unsure, I think that will eventually help to solve this brittleness problem.
Thanks for sharing all your practical insights with me!
Author info
