Welcome everyone to my talk on agent ai architecting, scalable and reliable AI agents for the future. This is Bitler working as a senior software development engineer at Adobe. Before I get on, get onto my talk, I want to make a quick disclaimer. All views expressed here are my own and do not reflect the opinions of any affiliated organization today. We. Tools to become active participants in solving complex problems. By the end, you'll understand the key architectural principles for building AI agents that are both scalable and reliable. Let's deep dive into it. If you really look at historically, AI has fundamentally transformed over the past decade. If you look at it, we started with systems that simply responded to queries, think early, search engine or basic chatbots. Then we moved to more sophisticated interactions. That the systems that could generate content or solve specific problems when prompted right, and now we're entering the era of agentic AI systems that can take initiatives, make decisions, and execute multi-step plans. This represents a fundamental shift in human computer interaction. From tools we used to partners we collaborate with. So that is the power of evolution that is happening, and that is the transformation that we are going through To understand agent ai, first of all, we need to define what makes system a truly agent. If we look at it, there are four qualities that I categorized here. A systems to act as an agent. Ai, the first goal is directed behavior. The system should have the ability to pursue specific objectives rather than just responding to immediate inputs. That is the first goal. And the second is. Operating with minimal human intervention. Once your goal is established, they should automatically perform all the tasks that are given. And the third one is environment awareness. The ability to perceive and interpret the context in which the agent operates and the adaptability. Learning from outcomes and adjusting strategies accordingly. So if these four capabilities are there in any system that we can say it as a agent system that can actually plug and play with any other bigger system, right? Let's look where already making an impact. There personal assistance, like advanced versions of tools that can proactively manage schedules for any individual filter information and coordinate their tasks with other agents are humans. In software development, we are seeing agents that can not just generate code, but plan implementations, debug issues, and performance. So in software development, we see in day to day, like you have, Microsoft Copilot, GitHub copilot, and also uc, cursor ai. These are amazing tools where you can actually leverage for code generation and couple more tools to perform. all of these tasks I mentioned, and if you look at the research. Design experiments and analyze the results. So for research agents, you have so many other types of things to take care. So now agents are there to solve all of these problems and in physical domains, robots, which are going to help you in industrial systems that can adapt to changing conditions and requirements. So that's where you can see the robotic agents in the manufacturing and in other support areas. So here, what's notable is the increasing scope of agency from narrow specific tasks to broader domains of action, right? So adaptability. willing to collaborate with the agent AI to perform their multitasking things without any stress, right? So now let's break down the essential components that enable the agent systems, right? So if you really look at the system here at the core, in the foundation. Typically large language models, which are going to provide a cognitive engine to break down the prompts and give the instructions. Then the planning frameworks comes into the picture for strategic decision making where the ability to map out the steps towards a goal to achieve the goal that is defined for the agents to integration systems. That can extend the agent's capabilities through APIs and external services. So that's where the integration comes into the picture. Then the memory systems for maintaining the context and history beyond immediate interaction. So see the storage related information comes where it can, where the agents can retrieve the information based on the. The system is able to learn from the success or failure and then try to, take the corresponding actions and the feedback mechanism. Also make the system so get corrected, process dynamically and stores the, information for the future usage. These are the typical key components of an agent system. So now. If you really look at this whole diagram in the centralized area where you have the total, the processing capability, like how the processing is going to be taking place, and there you have the lms, the large language models, which is the core of the system. And then these, there are multiple components. You see the bidirectional mapping between all of these tasks, right? That includes the planning and reasoning module where you see, how LLM need to, communicate within a different kind of, sub components and tool connectors, right? The a p connectors are the tools that can actually interact with between these, components. And the memory storage, right? So where you can store the outputs of your agent based on the kind of, information that is processed, right? And based on the historical data that it can actually keep it in, respond back. And then you also have monitoring and feedback where you have the feedback loops to see whether the system is, giving you the. And then the safety guardrails where you see like how the systems are actually, safe to use or not. So all of this information you can see through this technical architectural diagram. And then the challenges in scaling agentic ai, right? So as we move from prototype to production, there are several challenges. Increase dramatically with both model size and task complexity, then the latency becomes very critical. Users expect rapid responses even for complex tasks. So if you really look at the users the way they are coming up to the system, right? So you need to understand like how the agent is actually supporting what kind of complex. And then you see the reliability. Reliability, it must be maintained across an expanding set of use cases and age cases. And the multi-agent systems introduce new complexities in coordination and communication. And of course cost scales with huge, usage. And so these are. Are possible, but they require through the architectural approaches of agent AI and reliability concerns, reliability deserves a special attention as it's often the make or big factor for adoption. Because if without reliable systems, it's very difficult to, use it. So Hall. When agents generate convincing, but incorrect information remain a significant challenge, right? So if every information is coming as hallucination, it is very hard to soak that information and you can't really use it in the real time. Rule misuse occurs when agents apply available capabilities inappropriately or inefficiency. So that's where, to see where the tool is getting misused. Planning failures happen when agents create ineffective or nonsensical action sequences. So this is, planning is one of the crucial point. Context limitations. Context limitations lead to agents losing track of relevant information during extended tasks. So these are the major concerns, and apart from that, the last and final one is the feedback loops, which can trap agents in unpredictable patterns of behavior. So if the feedback loops are not giving you the proper feedback, if there is a chances of 20% errors information, and the feedback is actually going as a storage, eventually the system can give very information to the end user. So that's where you need to be very careful on what kind of feedback mechanism is there and what kind of corrections are getting. So here we, we have three solutions, so I want to go through one after the other. The fastest solution here is modular architecture, so by decomposing complex. We reduce the cognitive load on any single system, right? You're breaking down into multiple components and that's where you are, trying to give a task for each of the component in an isolation phase. This enables specialized agents for specific domains, leveraging expertise where it matters the most. Modularity facilitates easier updating of individual components without disrupting the entire system, and it supports distributed processing, spreading computational, and load across the resources. Perhaps most importantly, it allows for graceful degradation. If one component fails, other can continue functioning it. This approach has prevent successful in other complex software domains. So here, if you look at any kind of, banking system, how you can break down into multiple, agents and each agent can take care of its task. And ultimately they align with the centralized system to give the end results. And second solution is robust planning and framework. In this case, there is a hierarchical planning, which where the structure is break down into complex goals and into know manageable sub goals. So the ultimate goal is to, provide the solution to an end user with so many, interactive kind of processing that are needed for agents to work on it. So in this case, you may need to have verifications at multiple stages. Ensure that plans are logical and aligned with the common goal or the objective that is defined. In this case, you may see some uncertainty, estimation, uncertainty, which can help the agents to recognize when they may be operating with incomplete information. So this is where you may need to see like what is going wrong and which kind of scenario. So that's where the fallback mechanism provides. What are the alternate you have when the primary approach fails? So that's where you see here, this may be the common goal, but these agents are connected and everywhere you have a fallback mechanism where there can been corrective action, can be taken care, right? So again, when something goes wrong, there will be a dynamic planning. So the adaptable planning, Together, these approaches create more resilient planning for handling real world complexity. The third key solution in comprehensive, in comprehensive evaluation is it is like how you wanna test the quality of your agent. So in this case, you may have several test. However, not just common cases, but cases where failures are more likely, right? So continuous monitoring, which will allow us to detect the performance degradation before it becomes critical. So you may need to monitor, like whether is there any kind of underlying system is, going through any, performance bottlenecks or degradations. Real time failure detection enables immediate intervention when necessary, and also human feedback in cooperation ensures that technical metrics align with actual user needs. So human feedback is very essential because you can't really rely on a hundred percentage of the systems be. So this is what in the third solution, which you may need to, see like how it actually helps the, real user. So combining these three solutions, definitely you may get a better agent to use it as part of your implementation. So let's start with a case study. So here an example of a bank manager. So if you really look at an enterprise knowledge agent, how this works on the left hand side of this image, you can see in the current scenario how the bank manager is actually, working with the, in the real world scenario and on the right hand side, a knowledge agent that is enabling or using it for processing their tasks. Excuse me. So here, yeah, he's dealing with a lot of, started documents, databases, and systems. So he doesn't have any kind of a proper mechanism to handle the, current scenario here. This. This solution definitely a kind of a complex thing to really, handle, and he'll be stressful to, at the end of the day to handle, so many tasks. So that's where he got the knowledge agent in place. Okay. So this knowledge agent, what does it do for him? the solution implemented is a modular approach, where you got the, dis from different systems. And generation. So this way he is gonna get a scalable agent system with specialized components, which can be, implemented for documents. documents in a sense for reading, for writing, for any such kind of matters, any interpretation, like what kind of information is coming and how we need to process. And. Agent so that way a bank manager can relax and let the agents do the task for him and he can take the decisions as needed. The results, if you really look at both of these systems, these are, these results were remarkable. You can see 40% reduction information, re time. He doesn't need to really go personally, check the systems for report generation and all, and 65% improvement in accuracy. So there are no human errors and there is no kind of rush kind of a thing. So that way you can see, the systems are giving, like the improvement in accuracy and the reduction in information retrieval, right? This way people can relax and another case study is autonomous. The challenge. Where you can have, like with very minimal human oversight while maintaining their quality. So that is the objective. So if you wanna implement a solution to cater this need here, you can see like you can implement a agent system, which can perform a different roles. One agent could be for coding, another agent could be for testing, right? And another agent could be for integrating. Of course, integration testing. So if you use these agents, the results demonstrated out of this system would be a three x developer productivity and 50% reduction in bug. So developer can actually give quality, outcome, but using these agents and you can also be more efficient, right? So this is the outcome of agent, usage, especially in the autonomous software. General purpose agents. So these are like general purpose agents need a lot more information, lot more storage, lot more, processing capabilities, right? So here these move beyond narrow, specialized to handle diverse tasks across different domains. They maintain long memory and accumulate experience over extended. So they should have all the sophisticated information to maintain and to serve the clients. They transfer learning across domains and applying insights from one area to another. The domain includes, it could be banking, to insurance, and banking to, The trading systems, it can be of anything. They engage in meta reasoning about their own capabilities and limitations, and they also develop sophisticated understanding of human intent and beyond the explicit instructions. So these agents in the real world can solve a lot of problems where a common human can, struggle with. While we're not there yet, especially for the general purpose agents, architectural editions today will determine how quickly and safely we approach this horizon. Let's wait and see how soon that we can get there. But, and if you really look at, the agent AI or any a system for that matter, ethical considerations are very values. And ethical considerations. So how we can actually balance right? What is more important? So especially when it comes to the age and decision making, it's the transparency which is very essential for an appropriate oversight. And then the accountability mechanism we should ensure. Which should be built into a data usage and memory systems, right? You can't really use data for any reason. So the user data should be protected, and we need robust safeguard to prevent harmful emergent behaviors. So these are the things you may need to consider when you're. in successful systems, you need to see what can be incorporated, what shouldn't be. So separation of reasoning and action allows for specialized optimization of each function. And you should also have explicit verification steps, which should prevent cascading errors. And you should also have strategic human in bottleneck. If you use agents for everything, if agents stops doing mistakes, so if every time, if they're making 20% mistakes with five repetitions, you may see only the faulty system. So that's where human in the loop checkpoints always needed and graceful handling of uncertainty, which will avoid the. So anything that comes uncertainty, right? You should have always have the, modes to correct the system. And continuous learning from operational data, right? Which will be an ongoing improvement. So you should always incorporate that, right? So these are the imaging patterns you can use and directions. So if you really look at it as tasks become more complex, specialization becomes increasingly valuable. This requires a sophisticated communication, sorry, sophisticated communication protocols between the heterogeneous agents. so how these agents can communicate. a bank manager to process, currents and current account details and savings and loan account related information. So you may have the agents to process all of these information from different, distributed agents and giving that information to the, the bank manager to ensure that how the specific tasks are being performed. So you have the centralized information processing. Capability. And at the same time, agents are actually interacting with the central system and they're actually informing the bank manager to, use that information in an effective way. So this is the distributed, multi-agent, system that you can build. And ultimately one information will float outside. So in this case, the resource sharing and allocation mechanism becomes for efficiency. So they need to, like what kind of resources they need to share, like GPUs and memory or whatever it's, and conflict resolution. That is very important for any kind of agentic system to ensure that they are, prevented from deadlocks and any kind of contradictory actions. So they have to do, they should actually generate. Here, the ultimate goal is to emergent collective intelligence systems that can solve problems no single agent could solve with these kind of issues. So that's where you need to have the, collective agents to, process the information, future directions and adaptive systems. So these can dynamically adjust the capabilities based on the task requirements. And they improve themselves through operational excellence and not just leaving from data, but learning from their own mistakes. So this is the beauty of the adaptive systems. They automatically detect and mitigate weakness without human intervention. They may manage resources based on environmental constraints and opportunities. They implement context sensitive safety mechanisms that, that adapt to changing risk profiles. The result is system that become more capable and renewable over time through their own experience. So these are the systems you may want to see in the futuristic way, right? the self-learning systems are. Better. But at the same time, there should be checkpoints to ensure that, as I discussed in my previous slides, the human feedback loop mechanism we should definitely incorporate. But the self adaptable systems, you don't need to sit there and then train and they can learn by themselves and become more efficient. But at the same time, human feedback loops also, we should. Building for the future and the key principles for agent ai. To summarize the architectural approach, I advocate first, embrace modular design. So how do you wanna break down the whole systems into sub components? Second is prioritize the observability. So you should ensure that the systems are. Debug, if there is any kind of a system comes, system failure comes into the picture. Third is implement the controlled autonomy with clear boundaries for agent decision. And the fourth one is develop the scalable evaluation, the test system under diverse conditions. because you don't want system to, so it should be tested in diverse conditions. Pass engineering and see if the system is still reliable enough to, answer the, the questions asked by the user. Finally, integrate feedback continuously from deployment to, to develop and, improvement related things. So you should have the, con consistent feedback mechanism on the. Agent A represents a fundamental shift in how we interact with technology. So this is essential, like how you want to interact. these are the, whatever we discussed so far, these things really makes, agents are like more, when you maintain all of these things in a systematic way. So if you really look at now we are moving from tools we operate to partner with, and to partner with the agents, and also we also collaborate with them to ensure that, we get the, productivity improvement and accuracy improvement. This transition request, thoughtful architecture that balances capacity with reliability. So this is very essential transitioning, like when. Human based. So you have to, maintain it. And success purely depends on finding the right equilibrium between autonomy and oversight. And finally, the future belongs to composable, adaptable agent systems that can evolve with our needs. Thank you. Reach on.