We Built for Predictability; the Workloads Didn't Care#

Michael Stahnke, VP of Engineering at Flox (previously CircleCI and Puppet leadership)

• Why: Not networking, but it names a fault line that will run through DC and network operations: AI/agentic workloads are probabilistic, and config-management-era guarantees (determinism, no drift, reproducible environments) don't transfer to the application layer. For anyone whose products promise predictable infrastructure, this reframes what you're actually selling.
• What: An argument that infrastructure's predictability model doesn't fail but is simply insufficient for probabilistic workloads — a perfectly reproducible environment can still behave differently run to run — and a framework for deciding what must stay deterministic versus where randomness is safe (from abstract only).
• How: Conceptual/operational reasoning rather than a mechanism: partitioning systems into deterministic foundations and probabilistic workloads, and adapting troubleshooting and reliability practices for "it depends" outcomes.
• Where applicable: General and early — this is a mental-model talk, not a tool or result. Relevant as strategic input to product thinking about AI-era infrastructure requirements.