At QCon AI NYC 2025, Aaron Erickson emphasized that agentic AI is primarily an engineering challenge rather than just a prompt crafting task. He proposed that reliability stems from integrating probabilistic elements with deterministic frameworks. Erickson noted that agentic AI thrives when layered over existing operational systems, enabling question interpretation, evidence retrieval, and action suggestion, while deterministic systems handle execution and constraints.
He warned against the common pitfalls in language-to-SQL queries, highlighting that complexity leads to accuracy drops. His solution involved reducing variability through schema flattening and constrained queries. Additionally, he differentiated between classification and code generation, advocating for systems that route tasks to specialized agents to improve reliability.
Erickson also introduced a taxonomy of agent behaviors and championed role specialization, emphasizing that deterministic runbooks are essential for repeatability. He concluded with the importance of defining the boundaries between certainty and discovery in agentic systems, suggesting this is where effective platform engineering occurs. Full talk access will be available on January 15.
Source link
