Demetrios [00:00:00]: Let's get rocking and rolling into it with the first talk of the day and we'll do a whole roundtable session after all these talks. So hopefully you all have some burning thoughts on your mind. Agent SAYS NEURO SYMBOLIC Systems I'm so stoked for this. All right, I'll hand it over to you.

Nirodya Pussadeniya [00:00:23]: Yep, thank you Demetrius. And good evening to everyone in CET and very sleepy. Good morning to us over here in IST. It's around 12:30am so technically it's tomorrow morning, right Already. So yeah, let's dive in. So people are mimicking the idea that future is agentic, claiming that agents will do it all. So but here's the catch. So how we define agents and how should we see them, that's where the things get interesting.

Nirodya Pussadeniya [00:00:59]: So today I'll take you beyond the surface level ideas and trying to give a kind of a new way to see about agents and dive into the architecture, potential and necessity of neurosymbolic systems in defining what agents truly are and what they can achieve. So before we dive in, let me quickly introduce myself. So I'm Niroddi Prasadini and you can call me Ryan also. So I'm the co founder of Sineka and we focus on building production grade and practical AI solutions or tools that addresses real world problem with precision and adaptability. So my work revolves around building intelligent systems that go beyond the surface. And so today I want to share why I believe neurosymbolic agents are crucial. So yeah, let's take a moment to set the stage with neurosymbolic system. So it's a kind of a concept that it's not a kind of a new trend but a natural evolution of AI.

Nirodya Pussadeniya [00:02:09]: Because to understand why they matter, think of the two dominant paradigms in AI, neural network and symbolic systems. So when it comes to neural networks, they are fantastic at recognizing patterns. So they process unstructured data like text, images or audio and uncover insights that humans could never scale. So for an example, a neural network can quickly identify a cat or a dog in a photo and even if the image is blurry or incomplete. So but here's the catch. So they are essentially statistical machines, right? So they often lack the ability to reason logically or it might lead to hallucinations or outputs that may look convincing but are factually or logically incorrect. So imagine kind of an AI suggesting that cats can fly because it's statistically possible in its data set. So that's a real problem for task requiring positions.

Nirodya Pussadeniya [00:03:16]: And then when we are coming into the symbolic system on the other hand, so they excel at logical reasoning. So give them a structured problem like calculating taxes or planning train route with clear rules and they will handle it flawlessly. So they are predictable and consistent because every decision follow explicit predefined logic. But there's a downside because they are rigid, so they can't handle the messiness of real world data. For an example like imagine kind of a symbolic system designed to manage customer support queries. So if a user submits a request like the gadget I bought is acting weird, what should I do? So the system might struggle to interpret acting weird because it doesn't match any predefined rules or categories. So without manual intervention to add new rules for handling vague language, the system would fail to respond effectively. So this highlights how symbolic system lack the flexibility needed to adapt to unstructured real world inputs.

Nirodya Pussadeniya [00:04:33]: So now is the best part. Imagine combining these two approaches like picture and AI assistant managing kind of a complex logistic operation. So it uses neural network to understand unstructured inputs like vague customer emails or spoken instructions extracting key details. But then it switches to a symbolic reasoning mode to plan the more efficient, most efficient delivery routes based on traffic, data, fuel, cost or maybe delivery deadlines. So the neural component brings flexibility and the symbolic component ensures logical consistency. So this blending of strengths directly addresses current challenges in Agent Ki, like ensuring logical reasoning while adapting to messy real world inputs. So I believe those neuro symbolic systems don't just fix problems, but they redefine how we approach them and improve the reasoning capabilities of the agents. So yeah, so what are agents? Actually I might pretty sure that all of you have heard about this.

Nirodya Pussadeniya [00:05:44]: But to set the stage I'll go through it quickly. So, so let's take a step back and look at how we arrived at the concept of agents. So at first we had LLMs trained on vast data sets and they were powerful but limited. They could only rely on their training data which became outdated quickly. So this led to the next step, RAG or retrieval augmented generation. So RAC systems improved things by combining LLMs with external data sources and enabling them to use updated or domain specific information. But while RAC systems is all some problems, they were still limited, right? Because they were essentially smarter lookups. So what came next? Like we needed systems that could go further, system that could reason, that could use tools and adapt to new context and retain information over time.

Nirodya Pussadeniya [00:06:39]: So this need gave rise to agents. And agents are system designed to pursue reason and act. So but here's a question for you like. So is an agent simply a kind of a tool that can that build upon an LLM that can call some tools and perform a task. So think about right? So would you trust such a system to handle complex or complex real world workflows? The my answer is personally no, because a kind of a real agent is so much more than just a tool wrapper. It's a dynamic system capable of reasoning and adapting. So I believe that unfortunately most of the time when people build agents using frameworks like Langgraph 2, AI, Autogen or whatever kind of similar framework. So they are limited to creating workflows where an agent LLM simply call tools with few iterations to perform a task.

Nirodya Pussadeniya [00:07:44]: So while this might work for simple task, it's far from what needed for production level systems actually. So true agents go beyond this. They can reason, they can adapt and they can learn over time. So they operate with logical consistency and act intelligently in complex real world environments. So that's why it's time to rethink what agents can be and push the boundaries of what they are capable of. So now let me take you through a couple of examples where agents have been successfully implemented as neurosymbolic system. So this will give you a kind of a. Kind of an idea to view AI agent as a kind of a neural symbolic system.

Nirodya Pussadeniya [00:08:33]: So this example demonstrate how integrating neural and symbolic captoches create system that can handle complex tasks with adaptability and logic. So first we go with SOP agent. So this was came in later December I guess. So this framework is built around the idea of encoding workflows into symbolic decision graphs. Here's a kind of an overview picture that I took from the paper and this symbolic decision graph which act as a kind of a structured map for guiding the agent's behavior. So the SOP agent uses a navigator to form standard operating procedures, what you call it sops. So that breakdown task step by step. For each step the system filters varied actions and uses a neural model to select the next action.

Nirodya Pussadeniya [00:09:33]: So for an instance, let's say like when given a task to pick up and place a book on a shelf, it listens to a decision graph to logically plan and execute each step. So the result will be a kind of a combination of logical reason and neural component adaptability that makes agent dynamic and efficient. And also another example tree of code. So like unlike SOP agent, so unlike SOB agent which relies on predefined workflows, tree of code focuses on solving complex reasoning tasks by generating and executing code end to End so it uses a tree structure to explore solutions, iteratively refining them based on execution feedback and do their task. So for an example, in a query like find nearby restaurant within 1km of Colombo or maybe San Francisco. So tree of code generates step by step code, evaluates results at each stage and improves its solution dynamically. So this makes it ideal for multi step high complexity task requiring both logical planning and flexibility. So I quickly went through this, but if you are more interested I can post those couple of papers in the slack so you can go through it as well.

Nirodya Pussadeniya [00:11:07]: So both this op agent and tree of code highlight something crucial. So I believe that graphs are the backbone of neuro symbolic system because graphs naturally bridges neural and symbolic components. So symbolic representation mean kind of nodes and edges explicitly encode logical relationships enabling structured reasoning. And when it comes to neural flexibility that means embedding models enrich these graphs, adding learned patterns and enabling generalization. So this kind of thing happens with the graph. And graph also allow multi agent calibration as well. For an example, let's say one agent might analyze data using a neural network. So another agent uses the graph to reason logically about the data.

Nirodya Pussadeniya [00:11:59]: Together they refine decisions, ensuring consistency and adaptability. So yeah, so next, how we design systems like this. For me it starts with understanding that you need a foundation where both neural and symbolic components can coexist. So graphs are a natural starting point. They provide the structure and logic symbolic system to thrive on. So while also being flexible enough to integrate neural insights. So from there it's about building layers. The first layer might be a kind of a neural network that can process raw unstructured data and turning it into a recognizable patterns.

Nirodya Pussadeniya [00:12:50]: So then the second layer might be symbolic, taking those patterns and organizing them into a structured framework like graph or decision tree. So finally there's a third layer that brings it all together, a neural refinement stage that validates and contextualize the results. So ensuring that the system isn't just consistent but also relevant to the task at hand. So and I believe that it doesn't stop there. The system work best when you can incorporate multi agent collaboration. So one agent might specialize in understanding natural language, another in logical planning, another another in execution. So like by so when they work together through a shared graph based framework, the possibilities are limitless. So this is what I find exciting about neurosymbolic system and what I believe like why, why we should give kind of a definition towards agents in the frame of neurosymbolic system.

Nirodya Pussadeniya [00:13:54]: So they are not just about solving problems, but they are about doing doing it in a way that combines adaptability, reasoning and structure. So yeah, that's what agents should aspire to be. I believe so yeah. Before I wrap up, let me quickly share a kind of a personal project that reflects everything we have talked about today. So I'm currently building kind of a browser usage agent from scratch. So you might ask why? Because there are a lot of many tools already available which are performing good. But I believe that personally I believe that truly understand and innovate with agents. Sometimes you need to reinvent the wheel and to truly master the system you have to break it down and rebuild it yourself.

Nirodya Pussadeniya [00:14:49]: So it's an experiment to explore how agent can analyze, reason and act dynamically in practical setting. So I post about it in my Instagram and Facebook as well. So you can find my links to the socials through the QR. And if you are like to give a kind of a better understanding. This project combines neural models to analyze user behavior, symbolic structure to organize workflows and feedback loops to refine actions dynamically. So it's an experiment so but one that's helped me push the boundaries of what agents can do in real world scenarios. So if you find this is interesting, if you are working on something similar, let's connect and you can find me on LinkedIn as well. Yeah, you can scan the QR and connect with me and I would love to hear your thoughts, exchange ideas or even collaborate on next big thing.

Nirodya Pussadeniya [00:15:47]: So yeah, thank you for your time today.