QConAI: Balancing Probabilistic and Deterministic Systems for Reliable Agentic AI

Balancing Probabilistic and Deterministic Systems for Reliable Agentic AI

At QCon AI NYC 2025, Aaron Erickson positioned agentic AI as a core engineering problem, shifting the focus from prompt engineering to robust system design. He argued that reliable agentic AI requires a blend of probabilistic models for exploration and deterministic systems for execution and control.

Treating agentic AI as a layer over existing operational systems, rather than a replacement, is crucial for achieving reliability; however, relying solely on large language models for complex tasks like SQL query generation leads to rapidly diminishing accuracy as schema complexity increases. This accuracy degradation can translate to significant operational costs and potential failures in production systems.

Key Insights

• Schema Complexity & Accuracy: Accuracy of LLM-generated SQL queries falls sharply with complex schemas and numerous joins.
• Classification vs. Code Generation: Classification tasks exhibit higher reliability than arbitrary code generation.
• Tool Choice Paradox: LLMs struggle with excessive tool options, leading to suboptimal or unsafe selections.

Working Example

(No code included in the source context)

Practical Applications

• Fraud Detection: A system using an agent to classify transactions as potentially fraudulent, then routing them to a deterministic rule-based system for final adjudication.
• Pitfall: Over-reliance on LLMs for complex decision-making without deterministic safeguards can lead to unpredictable and potentially harmful outcomes.

References:

• https://www.infoq.com/news/2025/12/qcon-nvidia-platform/