In this lesson, we will discuss the key factors driving AI agents and their evolution in the field. You will learn about the current limitations and will explore possible future directions for where AI agent technology is headed, and how these developments might shape the AI landscape. Let's learn about the key factors driving AI agents proliferation. The first is technological advancements. Specialized hardware improvements such as GPUs and storage computing, larger models and more complex agents to run Python drivers. Second, computational and algorithm innovation provides a foundation for agents capable of planning and acting autonomously. With a variety of very different new advancements in LLMs. Another factor is data availability. Right now, we have large diverse data sets that enable agents to learn complex patterns for decision-making through open data initiatives and large scale platforms. Web agents have these current challenges. At first, from the lack of uniformity. There's, like, diverse frameworks and implementations, but there's no ways to make them work together. And there's a fragmented ecosystem. Where again there's a lot of things going on in that critical system. But interoperability is really hard. What is needed is some sort of standardization for interoperability, scalability and collaboration. As AI agents become increasingly adept at navigating the web, it's crucial to assess their performance in environments, that mirror real-world conditions. Traditional evaluation methods often fall short in replicating the dynamic and unpredictable nature of actual interactions. You're seeing too many websites, and need deterministic environments to ensure consistent conditions for each evaluations, eliminating variability that could affect performance assessments. Limited insight into user behavior and agent performance in web navigation. Can be corrected by realistic interactions that emulate real-world challenges. Can be corrected by realistic interactions that emulate real-world challenges. Including loading errors high latency, and disruptive pop-ups, to test agents' resilience and adaptability. Finally, eval methods fall short in unpredictable web interactions and need comprehensive evaluation metrics. But just as both the final outcomes of the transactions and the ability to navigate complex web scenarios effectively. To address these challenges, we are building a new benchmark: Realistic Evaluation for Agents Leaderboard. REAL. The REAL benchmark addresses previous challenges by providing a deterministic yet realistic playground that models real-world browser experiences. This setup allows for consistent evaluation of AI agents abilities to perform tasks such as information retrieval and transaction operations, and establishes a baseline for browser agent models. This collective approach aims to establish robust baselines and leaderboards for state-of-the-art agents, benefiting both the public and academic sectors. I encourage you to engage with the project by providing feedback. Sharing insights and contributing to the refinement of evaluation scenarios. By collaborating, we can enhance the evaluation frameworks for AI agents, ensuring that they are well equipped to handle the complexities of the real world web navigation tasks. So far we've discussed a single AI agent. Now, as an user, you can have a multi-system for the manager agent that oversees multiple specialized worker agents. However, with everyone creating their own agents. This can lead to a chaotic ecosystem. Let's explore the challenges and considerations for Multi-Agent Systems. First one is the coordination complexity that is ensuring coherent behavior among autonomous agents, which can be challenging without a centralized oversight. Second, communication overhead. Decent resistance may require more complex communication protocols to facilitate effective collaboration. And third, security concerns. Maintaining system integrity and preventing malicious behavior, and an open, decentralized environment necessitates robust security measures. What is needed is modularity, specialization, and control. Here are some future research directions that you see. First, exploring hybrid architectures that combine centralized and decentralized elements to leverage the advantages on both approaches. Advanced communication protocols that are more effective and allow for better coordination. And third, investigating the application of decentralized multi-agent systems in different domains such as collaborative robotics, sensing and simulations. Here are some Multi-Agent Architectures that need further explorations. That is the senior supervisor. Using LLM for tool-calling. Having a network-based architecture. Having custom multi-agent workflows. And having hierarchy. Now, we take a look at the future trends and actions. Web agents advancing capabilities for better natural language understanding and tool use. The deep learning will enable for students to perform complex sequence of actions online. Such as appointments, researching and summarizing information coordinating across websites all on user's behalf. Multi-agent systems cooperate in a shared environment and learn sophisticated strategies by interacting with each other. Finally, Advances in distributed cloud computing will support the other advancements, and lead to building more scalable systems that can be deployed in the real world. In this lesson we explored some of the challenges in the current agentic system. And the future trends on how this can be standardized and improved for deployment.