From Prompt to Protocol: Designing Agent-Oriented Infrastructure for LLM Workflows

Large Language Models (LLMs) have evolved from passive text generators into active agents capable of perception, reasoning, and coordinated action. However, most current AI applications still rely on ad-hoc prompting rather than systematic architectures for multi-agent collaboration and workflow automation. This talk explores the transition from prompt engineering to protocol design, building structured, agent-oriented infrastructures that enable LLMs to communicate, plan, and execute tasks reliably across domains. We discuss emerging paradigms such as multi-agent orchestration, memory persistence, and self-correcting feedback loops that transform LLMs into components of dynamic, goal-driven systems. Through lightweight simulations and GPT-based experiments, we demonstrate how protocolized communication, using message schemas, roles, and verification stages, enhances task efficiency, reduces hallucination, and supports modular scaling. The session concludes by outlining design principles for AI systems that move beyond single-turn prompts toward continuous, autonomous workflows, bridging the gap between human intention and machine execution.

Shreya Singhal is an AI researcher and engineer specializing in data science, machine learning, and generative AI. She is pursuing her Master’s in Computer Science at the University of Texas at Austin and has led multiple end-to-end AI projects at Claritev, focusing on multimodal models, LLM fine-tuning, and structured data extraction. Her experience spans research and industry roles at Dell Technologies, Aristocrat Gaming, and UT Austin, where she has developed scalable AI agents, RAG pipelines, and bias analyses in word embeddings. Her research on multimodal humor generation has been published under the title “Joke Generation using Multimodals and Generative AI for Abstract Images.” Passionate about advancing responsible and creative AI, she works at the intersection of language, vision, and human-centered intelligence.