Large language models (LLMs) have matured enough to power autonomous AI agents that can understand nuanced context, operate various tools with minimal human intervention, and perform multi-step tasks. Unlike other generative AI applications, these agents actively pursue goals and decide on the right tools to achieve them. They combine several key mechanisms to tackle complex tasks effectively, including understanding a situation, weighing options, and choosing the best path forward. Agents use well-defined tools and APIs, clear instructions, memory, and continuous evaluation to accomplish specific tasks. The modular nature of compound AI systems enables teams to start simple and incrementally expand system capabilities, making practical decisions about system architecture as they evolve. Evaluating agents requires tracking outcomes and process efficiency, and the deepset AI Platform empowers users to harness this new class of AI by rapidly prototyping agentic solutions and iterating based on real-world feedback.