AI agents deployed for IT operations often represent permanent cost centers, as each execution necessitates full LLM inference, even for previously resolved issues. This paper introduces progressive crystallization, a lifecycle that perceives agent exploration as a discovery mechanism rather than a permanent execution model. It defines a three-stage execution taxonomy, ranging from fully agent-orchestrated to hybrid to fully deterministic workflows, alongside an evidence-based promotion mechanism that converts repeatedly validated agent behaviors into cheaper and more reproducible deterministic workflows while automatically demoting regressive workflows. Evaluated on a production cloud networking AIOps system processing tens of thousands of incidents monthly, the approach increased deterministic execution from 0% to 45% over eight months, reduced per-incident agent costs by over 70% despite a doubling in incident volume, and enhanced safety through greater reproducibility and auditability. The paper also discusses the execution taxonomy, promotion and demotion criteria, trace extraction methodology, economic model, safety considerations, and limitations and threats to validity.
Blogger's Review: This paper presents a significant reduction in operational costs for AI agents while enhancing workflow determinism and auditability through progressive crystallization. It offers fresh insights into IT operations, especially in high-incident scenarios, showcasing how agent exploration can be effectively transformed into efficient production workflows, making it a noteworthy contribution to the field.