Building LLM-Powered Agents with Real-Time Reasoning Loops - Apurva Misra

This talk explores the evolution of generative AI systems from single-shot prompts to closed-loop reasoning, focusing on the structured processes of planning, acting, observing, reflecting, and refining. The speaker provides a production-focused perspective on building real-time reasoning agents from first principles, detailing how to integrate planning, memory, and tool use into adaptive reasoning loops that enhance agent reliability. Participants learn about frameworks like BAML and PydanticAI, as well as techniques for tool calling, memory management, and implementing guardrails. A live demonstration showcases an agent that effectively plans research steps, utilizes tools, and generates structured outputs in real time.