What stays constant
Which stages remain stable across classical ML, retrieval-heavy workflows, prompt systems, and agentic systems?
AI Operating Model
This page is for the durable shape of the work.
The working assumption here is that classical MLOps and newer context-heavy GenAI work still share a lifecycle, even when the tooling looks different.
What this page should eventually answer
Section titled “What this page should eventually answer”What actually changes
Where do prompts, context, tools, evaluations, and provider constraints change the operating model in ways that matter?
Lifecycle slices worth documenting
Section titled “Lifecycle slices worth documenting”Intent and framing
Problem definition, users, risks, and what “good” means.
Data and context
Source material, retrieval surfaces, freshness, and trust boundaries.
Build and evaluation
Experiment loops, evals, regression checks, and review gates.
Release and operations
Deployment, observability, rollback, and ongoing governance.
Questions worth coming back to
Section titled “Questions worth coming back to”System questions
- Which artifacts are versioned, and which are still treated as loose operational state?
- Where do evaluation results become release gates rather than just diagnostics?
- How do provider dependencies change the deployment and rollback model?
- Which decisions live with the platform team versus the product team?