A 37 minute presentation and discussion on TrueAGENT. Our Agentic Framework webinar is going to be interactive so please ask questions as we go. There is a Q&A button. We will be demoing the benefits and advantages of adopting an agent-based AI solution. Alongside that, the framework that can bring significant business benefits.
There are going to be three of us involved from AI on Cloud: There will be some live demonstrations as well, two live demonstrations giving you two examples of how the framework can help you deliver efficiently and with rapid demonstrable benefits.
Myself, the CRO, the moderator and general scene setter for this session. Nikos, the managing partner of AI on Cloud, is going to provide the principles of agents and true agent and its benefits and guide our CTO and the lead developer of this product.
Transcript
Good afternoon. Welcome to our Agentic Framework webinar. It’s going to be interactive. Please ask questions as we go. There is a Q&A button. We will be demoing the benefits and advantages of adopting an agent-based AI solution. And alongside that, the framework that can bring significant business benefits.
There are going to be three of us involved from AI on Cloud:
- Myself, the CRO of AI on Cloud, the moderator and general scene setter for this session.
- Nikos, the managing partner of AI on Cloud, is going to provide the principles of agents and true agent and its benefits and guide our CTO and the lead developer of this product.
- There will be some live demonstrations as well, two live demonstrations giving you two examples of how the framework can help you deliver efficiently and with rapid demonstrable benefits.
Nikos, next slide please.
The agenda is a simple one. First, we’re going to just run through some of the business cases, our real-world experiences of the business cases and how the AI agents deliver. We’re going to explain somewhat the why and the how and the wherefore of true agent, both via slideware and live demonstrations. And then finally, we will go into any next steps, answer your questions, etc.
So really, I guess the question is, are you ready for autonomous AI?
Very quickly, and this is the only slide on AI on Cloud today, just to give you the high-level view of what we are. We’re a consultancy, very focused on business implementations of Gen AI. We help companies, customers develop roadmaps, suggest the ripest areas of your business for AI implementation, help you proof of concept them and implement them.
As a result of some of these projects, we’ve also developed two product stroke solutions. The first, TrueRAG, we held a webinar on a few weeks ago. If you have any business requirements where you would be really keen on eliminating hallucinations and always getting the right answer, the truthful answer from AI, please go to our website where you’ll find a recording of last month’s webinar. But today, let’s focus on TrueAGENT, our agentic framework.
Next slide, Nikos. Next slide, please. Thank you.
So on their own, technologies such as automation, orchestration, AI, APIs, can achieve a lot. However, by combining these into one cohesive framework, you can generate solutions that work seamlessly together, creating more impactful, more relevant outcomes across your business.
The frameworks are effectively systems designed to enable AI agents to perform tasks with greater autonomy, taking away some of the human, but not all of the human interfaces. Maybe in the future, it would take away all of the human elements. But these frameworks provide structure and mechanisms for AI solutions, systems to understand the goals and the tasks to achieve those goals, plan a sequence of actions, execute those actions, observe, audit, record those outcomes, and then learn so you go back around and cycle through those results.
As you can imagine, this sort of implementation of AI delivers business benefits across a wide range of areas. And I’m not planning on listing or going through all of them. You can probably imagine many of them, but:
- Enhance productivity and efficiency.
- The automation of end-to-end tasks with minimal human intervention, freeing staff for higher value, more customer-facing work.
- Going to reduce completion time for workflows by eliminating handoffs.
- And maybe most dramatically at times, operate continuously without breaks. AI agents just don’t need breaks.
Overall, this is going to reduce your operational costs and improve your accuracy. Other areas where it canwork, not quite yet, you’ve jumped a bit quick there, Nikos. We really need to get an AI agent that presses the button when I tell them to. Greater adaptability and resilience. You can respond to changing conditions dynamically. One of the demonstrations you’re going to see is a demonstration where effectively the agent has learned through reading and understanding the user manual. If you update the user manual, the AI agent will, an application will change and learn and change automatically. This accelerates innovation, brings you competitive advantage. If your competitors are doing this, you will beat them up through traditional automation. Create systems that improve over time through autonomous learning. And you will eventually, we will eventually get to where for many mundane, repetitive tasks, AI agents will be autonomous.
Now we’ll try the next one, Nikos. But why does this matter? Well, rather than just believe me, just want to make sure that you don’t think it’s just me saying this. McKinsey in second half of 2024 did a review of Gen AI adoption. By the end of 2025, that’s this year, they expect something around:
- 20% of C-suite executives to be using AI to do approximately a third of their work.
- 50% of employees to be using AI to do approximately a third of their work.
And after that, through one to five years and over five years, it just continues to accelerate. Anyway, enough scene setting. To expand on this, I’m going to hand over to Nikos and then later on to Guy to delve deeper into AI agents. Over to you, Nikos.
Thank you very much, Neil. So I think to delve into how do we actually achieve this automation and offer the benefits to the business that you promised, Neil, it’s good to see what are actually Gen AI agents and agenting frameworks and where does true agent as a solution play in. So a lot of people would have used Gen AI so far through the typical interactive query and answer basis where you go into OpenAI’s ChatGPT UI or you go into any of the other vendors through their UI and you prompt, you even put some documents in or some other information and the LLM reviews that and gives you an answer. And it could be text-based, video-based, audio-based, but it’s something that you give, you get an answer and then you act on the response. And this is the traditional way. This is the way we’ve learned how to use AI so far. And a lot of people think about AI as being just that, something that you give a question and you get some answer. And if you’re lucky, you might have given it additional information in the form of some knowledge base that it has used to be more intelligent on how it’s answering back to you and guiding you.
But if you were to go a step further and wrap the LLM with what you’d call flow control software and allow to use its prompt to determine the next best action, then you can get it to effectively do what you would be doing by receiving its answer back. And that’s when we start talking about agentic frameworks. So basically what has happened here is we’ve taken the LLM, we’ve wrapped it with control software that allows the agent itself, based on its answers, triggered by the question you’ve answered or the request you put in and potentially using some information, determine what it should be doing on your behalf to take the next action and do what you are meant to do.
So in essence, if we expand that out, what you actually can achieve taking this approach is a situation where the agent, which is now turning into a more intelligent piece of software, where the LLM is not guiding you, but it’s guidingitself, can take some input and take some actions that’s been guided against through its processing what you told it and using its output as further input to do what you would have done instead and give you the result that you would have had to do work for to get to. And in this process, you can also give it some tools that could be tools that you would have been using, but you actually allow it to use to interact potentially with applications or do things that you would have done otherwise to get to these results.
So effectively, the agent and the tools are in essence replacing a user in taking actions. And the only thing you have to do then is to prompt it to do what it needs to do and it’ll come up with a result. So you can think of the agents as software robots in essence that use LLM as a brain. And by using the tools, actually the brain has the hands and the tools to come to the output.
Imagine a situation where, you know, and why you need the tools, you need the tools because if you let the agent on its own do things, you might get, you know, depending on how clever the agent is, might be able to do or not do things. Or it might be very specific actions that you need to take that the agent without the tool cannot take. For example, interfacing to an ERP application. The agent would probably not know how that ERP application works. You might have to give it some API or something to work with that ERP solution to do what is required to be doing, the same way that a user would have been doing the same thing.
So on this basis, we would like to take you then to some example usage of what this would look like in real life and also use that as a basis for the demos that Guy is going to show you in a few minutes.
So we’re going to look at the agentic usage on two scenarios:
- A business data analytics process, business process.
- Executing a business process by using a legacy application.
In the first instance, if somebody was to ask a business analyst to create some kind of report, that business analyst will have to go into a relational database, create some SQL, retrieve some data, and pass the data through some kind of reporting tool to present some report to its management that made sense for business action.
In the second case, you would have a user relying on using a legacy application, the UAF legacy application, to execute a business process. And what we’re going to show you is how an AI agent, how our solution through agent can do that.
So as I said in the first case, and the best that the business analyst can hope for in the traditional way of using UI is to use an LLM to ask questions about how to create the SQL and how to manipulate the data, what tools should be using, how he should do the work, but the person, the user is in the middle of the whole process, is the person that he has to ask the LLM, take the output SQL query, run it against the database, get the data, format it, and analyze it, and then present the report.
In our case, by using through agent, what happens is actually very automatic. Using our solution that is based on the agenting framework principles, what happens is that the user asks the through agent environment solution to create the graph, and it asks it, this is what I want, and the through agent solution goes to the SQL, creates the right SQL because it knows the schema, creates the right SQL, retrieves the data, does the analysis, creates the graph, and provides the user with the output. Therefore, the user becomes the customer now, and the agent becomes what the user used to do. And you can imagine that the time to production is a lot shorter, and the
The user of a legacy application, the best they can hope for in the traditional use of an AI is to send the user manual or the quality is also improved because we have a stronger brain, in a sense, in the form of an LLM doing what needs to be done.
questions to an LLM and get it to provide back to the user a simplified version of instructions of how to use the application, assuming there is a user manual. We assume there is because obviously somehow that user needs to know how to use the application. So that’s the best we can hope, is send the user manual and hope that we get good enough instructions back to interact with the legacy application in a more intelligent way.
In our case, and I think this is where our solution really stands out compared to what there is in the market at the moment, the agent, the through agent solution, is actually going to interact with the legacy application on behalf of the user using a local agent, a local application or a utility that the only thing that it knows is how to drive the UI. It knows how to press buttons and move the mouse, nothing else. It knows nothing about the legacy application and gets all the requests from the agent. So the agent is telling it, as the user would have done in the past:
- Click this button,
- Move there,
- Put that data into this field, etc.
to process the user request from A to Z, do what the user would have done by clicking the same buttons and filling up the same fields in the screen. And in this case you can trigger this process through an email or through the web browser or whatever. In our case it will be an email going in and the agent using the right tool to receive the request from the user and then the execution tool that basically has read the user manual in essence of the application, knows how the application works, has also received the request from the user, has put it together and then knows that it needs to start the application, go into specific screens, press the right buttons and get the results, get the output required and effectively close the application as well at the end of the day.
And the other important thing here is that in our implementation, in our solution, the local agent creates an outbound process, a connection to the true agent. So this connection here is an outbound one which is totally secure and without this connection the AI agent wouldn’t know where to send the request. So this is a totally secure environment, it’s a watertight environment that nobody can access from the outside unless the local agent is connected. Which means that if this legacy application is sitting in a totally secure environment, we will not be disturbing that kind of security. Our implementation is absolutely fitting into very sensitive secure environments that need to be driven.
And having said all that about our true agent as a solution, in essence just delving a little bit under the hood, it’s a solution rather than the framework. So it’s not one of the traditional frameworks you have sitting around with screw AI or LangGraph etc. where you have to actually do all of the coding to get somewhere. Ours is a solution, so therefore minimum coding required. Perhaps the only thing you have to give it is the appropriate tools, if that, because we have quite a few of the tools there, especially in this solution which is a critical one where it’s enabling automation. And that legacy application might, you know, it does need to be legacy, it could be any application, right? It could be an API, it could be a modern application, it could be an API to some, you know, whatever application you want to drive. So we can drive any application on behalf of the user.
It is a solution that’s based on cloud deployment, so you get by default all the cloud benefits of security, scalability, availability, observability and logging out of the box. And, you know, if you trust an application on the cloud, you’ll trust our solution. It’s plug and play in terms of tool deployment. And the interesting thing here, in contrast to all the other solutions,engineering framework solutions that are sitting, that are around at the moment is that the other solutions would need you to either use Python or JavaScript. With ours, you can use any of the modern languages that are listed here plus to create new tools so you can reuse either existing tooling that you have, repurpose it, or you can write it in a language that everybody knows rather than just have to learn Python or JavaScript or TypeScript.
And of course, on top of that, it’s a solution, it’s not just a framework. The other thing is obviously you can use any model, any AI model to be driven. It doesn’t need to be the ones that the cloud provider is allowing you to use or what the other main frameworks in the market are using because they are frameworks and solutions. In ours, it’s designed from the beginning to be able to use any of the modern LLMs that are coming to the market.
And finally, it’s a pay-as-you-go cost model. So there’s no upfront fee or there’s no lock-in with a fee that is standard. You pay usually, it’s exactly as the cloud principles of pay-as-you-go. So if you use it, you obviously pay more, but if you don’t use it as much, it’s not costing you money, burning money out of your budget.
So this is the TrueAGENT as a solution in a nutshell. And I would like to hand it over now to Guy to run us through the first demo. On the second demo, we would have to interact with Guy, so I will come back on that.
So Guy, over to you, take the screen. Thank you, Nikos, for the introduction. So I will start with the first demo of the data analytics. I will use some public data set. I’m calling from the US, so here baseball is the thing. So I can show it on a baseball data set. I hope that you can translate this example to any SQL interface, any SQL data that you might have in your organization.
So in this screen, I can choose which agent I want to use. I will choose the SQL agent. It can be interesting, I will show you in a second. We’re using Claude from Anthropik. But we can, as Nico mentioned, we can switch it to Gemini, to GPT, to DeepSeek, whatever is the LLM that you prefer for the task or budget-wise.
And I can ask a question. Again, I will take some kind of baseball-related question. The nice thing about it is using natural language. The business user, it doesn’t have to be a business analyst, it doesn’t need to know how SQL is written, what is the structure of a database. I ask the question, and now the agent is running.
I will jump behind the scenes so you can see more or less what is happening. I opened the hood for you. So this is the agent. You see that it is running. I can have the observability of what is the structure of the agent, what tools it has, what it is doing in each one of the steps. So now it is executing a SQL query. So you can see that it is thinking. And we’ll see in a second what are the steps that it did to get the answer.
Again, thinking again, now it is executing some kind of a Python code to be able to generate the graph that we’re going to see in a second. Now it’s doing another query. So the agent is basically thinking, how can I answer the question, the business question that I was posed, by using internal tools that I have in my organization. It will call the server database, the CRM, whatever type of data you have internally.
Still thinking, building another graph. And now when we go back to our interface, I hope that we’ll get our answer. So it said, OK, I was successful. And then you can also see the thought process that it did. So the first thing it said, I need to know what is the scheme of the database. I cannot create a query without that. So this is one table, and there’s another table with many, many columns. They will easily understand that. Now it’s starting to generatesome queries to get the data that it needs. So first it’s checking for the salaries of the people, and then generating some graphs. And this is the first graph it generated. So the top high paid 2005 were this player. Really nice visualization. So no need to build any dashboard. It’s basically thinking how can I help you understand that and generate all the information that is needed in a very user-friendly interface. Again, this interface that we show here is only for debug purposes. You don’t need to see all those back and forth of the agent. We show it for you to understand how the AI is thinking. But the user will basically get, these are the reply, these are the graphs, what is your next question?
So this is the first example of the SQL, the data analytics example. If you have questions on what kind of database we support, what kind of, is it only SQL, can we do other queries, all those questions we can talk about. Please use the Q&A window for what is it you care about. And we can jump to the second demo, which is again much more complicated. As Nico mentioned, it’s unique. It’s very hard to do that. We saw many companies trying to do that with one way or another. And our solution, I think, makes a bit more sense.
So Nico, you want to switch to your legacy Windows behind the firewall machine? Yes. Thank you. So just very quickly before I do so, just to remind people what you’re going to see. So what we’re going to see is effectively a request going in via email. And we’re going to see the local agent effectively driven by the LLM to drive the legacy application.
So if I can switch over to my legacy machine, which is actually my desktop at home, that cannot be seen from the internet. You could not access this actually because sitting behind a fiber optic router and nobody can access it, even if you knew my IP. What you see here basically is the UI driver, I call it this, waiting to receive commands from TrueAGENT and do things to the application.
And what I’m going to say to Guy, presuming that Guy’s inbox has been monitored by TrueAGENT, and I’m going to send this request, which is to update the affected day of an account, account 0101 to 24 or 6 of 25. Lovely date because the day of my birthday. So, you know, update the account that day when you receive this email.
And what the agent, what TrueAGENT is going to do is going to intercept this email and take the appropriate actions for this to go into, to be in effect. And Nikos, before you press send, just to make sure that people feel comfortable that it’s, this is not rigged. Maybe change the date from 24th of June to my birthday.
Your birthday. Definitely can do that. No problem. What is your date? 3rd of August. 3rd of August. All right. It’s coming in a while, so don’t forget. Sorry, 3rd of August. Sorry. Okay. Don’t be obliged by the present. Not yet. I will remember. Now hold your, press send and hold your breath. Yeah, and hold the hands up. Okay, I will do so. So I’m holding my hands up. Nothing is, I’m not touching anything on my keyboard anymore. The agent is monitoring regularly the inbox and the minute it will pick up the email, it will analyze it and try to come up with what is the best plan to implement this request. Once it has this plan, it will send it, put it as an activity and the local agent should put it.
I see it’s reading the emails here on the back end. I will share with you at the end the back end view. Thinking, thinking. Hopefully we won’t hit by the curse of the real time. There you go. So it’s openedthe application. It typed what this type, it searched for it, it found it. It typed in the date, which is different. Thank you very much. I should have sent the original email. Maybe I should have opened it up to show that it wasn’t actually that date. It was a different date. For the eagle-eyed on you, you would have seen that the date was actually different when the application started.
Yeah, so this is basically… It was successful, and if I will share with you the backend view, so you already have some view on how do we run and how do we visualize the agent.
So the agent, as we said, we have an LLM. The LLM, we can edit it and change the LLM to be a different one, again, based on the performance, the cost, the security aspect, whatever is the reason, and we want one LLM and other. This is one limitation from other frameworks, especially from the cloud provider, that you can use only the best that they can offer, Gemini and Google, GPT in Azure, and Claude or Nova in AWS, and we give you a bit more freedom there.
And then what are the different tools you want to have as part of this agent, and you can assemble different tools for different tasks. We saw the SQL task before that we have this SQL generation, schema exploration, the code to generate the graphs, all those were part of the SQL.
And here we have something else, something that is basically using the Microsoft Graph API to read the emails, and then the remote execution. So you can assemble for each task what are the right tools for that.
And you saw also the local agent that knows how to interact securely with the agent by just pulling the script, the request from the agent. You see here this is the remote execution automation script part, that either we pick it up and it’s executed or maybe nobody picked it up or there was a failure in execution and we need to retry and do something else.
So this is the behind the scene, under the hood view of the agent. We can see observability of all the steps that we performed, how much time it took, if there was a failure we can go and retry the application or debug what is the issue.
So all those things that are usually very, very hard when you build an agent, the flexibility that we talked about, the observability that is important to make sure that it continues to run, all those are things that it’s quite hard to get and we spend a lot of effort based on the experience when working with quite a few customers that we try to use other tools and we saw that they lack this capability.
And thanks just very quickly if I may add two things here.
- The demo that we have shown is actually a redacted version of a real use case that we put out for deployed for a customer. So it’s not theoretical, it’s actually a real application that is running on behalf of a customer at the moment that we have deployed.
- And secondly, and I think that’s important, to extend the functionality to additional screens if the application was, if we were comfortable with doing automating part of the application and we want to expand to other flows through this application or indeed to bring new applications in, the delta is literally adding those pages of the user manual into the knowledge base that the solution can access to learn how to use the additional screens of the application or the new application to execute the user request.
We don’t have to go and analyze the screen specifically and do lots of detailed teaching of the agent on how to use the application in a sense.
And what AI is doing, it’s creating some kind of a diagram that you can see here how it’s represented inside of its memory to how do I switch from the edit screen to the search screen to the close screen and so on.
So all those learning, it’s very, very powerful to be able to do many, many different tasks and drive quite a few applications in a very flexible way. So it’s
Easy for you to teach the AI how to drive your application and doing it on legacy applications on Windows, web browsers, and any other things that you want.
to do. And we know that in every big organization, there are hundreds of those interfaces that you would never modernize or you modernize in a few years in the future. And until then, you have no option to automate those processes. And now I think that you have a very good option.
Guy, thank you very much. I say from a more business point of view, David mentioned it, but I just like to absolutely be crystal clear on this for everyone on the call. That is a legacy application that has no API. It is AI driving the mouse and the cursor to enter the correct information into the correct field. We’ve not had to teach it that the field is in the top right-hand corner or the bottom left-hand corner or wherever it may be.
So that brings us towards the end of our session. We’ve deliberately tried to keep it shortish. We’ve got 10 minutes or so left. Should there be any questions or thoughts? You’ve got one more slide that you want to do. I just leave it as it is. I think we covered all these areas before, but just leaving people with kind of a summary of what we said.
So basically, why TrueAGENT versus the agent frameworks that you can see there? First of all, because it’s a solution and it’s not just a framework. You can use any of the programming languages that people love and like and know how to use to create tools. You can use any LLM, not the ones that your cloud provider is restricting you or the agent frameworks that are at the market at the moment are restricting you. And they usually like to use OpenAI. We can use anything under the sun.
It’s a pay-as-you-go model. It has built-in security, scalability, availability, and observability because by the nature of being a cloud-based solution. So these are areas that you don’t have to worry about, whereas you will have to worry if you use a framework. It’s a de facto thing that if you use a cloud solution, all these things are baked in.
And these are the key differentiators between our approach, which is a solution, as I said, versus going out and using Crew AI or AutoGen or whatever have you in the market at the moment, which is you have to build from scratch. It’s the difference between going to IKEA and buying a table that needs some assembly required versus buying a table from a store that serves tables, right? This way, you still have to buy the plates and the forks and the knives.
Is there anyone out there that has got any questions or challenges for us? We have recorded this. We will send you a link to it. It will be published onto our website. Should you want to review it, if you come up with any questions, please do not hesitate to contact us. Other than that, we thank you very much for your time and have a great rest of the day.
