Vasuman discusses the complexities of building AI agents, emphasizing that they are not simple plug-and-play solutions. Many people misunderstand what an agent is, leading to failures when they try to implement them in real-world scenarios. He notes that while agents can perform effectively in demonstrations, they often break down in production due to issues like hallucination and lack of memory. Having spent over a year deploying agents after leaving Meta, Vasuman shares insights from his experience, highlighting that successful implementation comes from learning from failures. One of the key lessons is the importance of context. Agents need to retain relevant historical information, not just focus on the current task. For instance, an agent managing invoice exceptions must remember past interactions and details about the vendor to make informed decisions. Poor context management can lead to agents making contradictory decisions or repeating unnecessary tasks. He stresses that effective agents can connect different pieces of information without explicit instructions, which is essential for producing results in enterprise settings. Vasuman also addresses how agents can multiply outcomes rather than simply replace human labor. He argues that agents can reduce the time humans spend on mundane tasks, allowing them to focus on more valuable decision-making. However, he points out that this doesn’t mean companies are laying off employees; instead, there’s still a significant amount of work that requires human oversight. The businesses that succeed with agents recognize that the real value lies in enhancing human capabilities rather than eliminating them. Lastly, he discusses how agents manage memory and state across tasks. He categorizes agents into three patterns: solo agents that handle entire workflows, parallel agents that tackle different parts of a problem, and collaborative agents that pass tasks between each other. Each pattern has its own challenges, especially in maintaining context and resolving conflicts. Understanding these architectural choices is critical for building effective agents that can scale and operate successfully in complex environments.