AI services need a catalog the same way every other internal platform does. The wiki approach falls over the moment you have more than a handful of models. Backstage with a thin AI plugin layer is the pattern that holds, a direct callback to the catalog discipline.
The interesting work in any AI system lives in the model and the application layer. The boring middleware between them, auth, rate limits, retries, logging, request shaping, is what makes the system actually work. Worth defending the boring part.
If your AI infra depends on third-party container images, you don't control your supply chain. Forgejo on store-01 as the source-of-truth git host, Harbor on engine-01 as the registry plus image-signing layer. The sovereign-infra argument, and why mirroring is non-negotiable now.
Not every AI workload belongs on Kubernetes. Some belong nowhere else. The patterns that hold up, separating CPU and GPU tiers, sizing autoscaling for serving versus batch, picking the right foundation, and the ones that fall apart at the first real load.
Cloud lock-in in 2010 was bad. AI lock-in in 2025 is worse for reasons most teams aren't thinking about. The data, the prompt patterns, the agentic surface, the fine-tunes, none of it ports cleanly. Worth being clear about why before you commit.
The right unit for AI cost tracking isn't tokens, isn't requests, isn't users. It's conversations. The cost story makes sense when you measure it that way and gets impossible when you don't.
Tekton, Argo CD, GitHub Actions, Jenkins X, four answers to model deploys. You can't unit-test a model, so eval suites become the test substitute. Versioning, rollback, blue-green serving. Pipeline config as the Decisions as Code surface, projected per environment.
The infrastructure-as-code review pipeline is one of the highest-leverage places to deploy AI in an engineering org, and almost nobody is doing it well. The reasons it's underused are mostly accidental rather than principled.
Every platform team buries the business in an eighty-nine-row YAML and calls it self-service. That isn't self-service, it's a configuration handoff. Decisions as Code is the discipline of extracting the five decisions that actually matter and letting the platform absorb the rest.
AWS shipped its own agent SDK two weeks ago, joining LangGraph, Semantic Kernel, the OpenAI Agents SDK, the Anthropic Agents SDK, and a half-dozen others. The question isn't which one's best. It's whether the framework layer is still where the differentiation happens at all.
Build was the agents conference Microsoft has been preparing for. The agents pitch lands. The governance story underneath it is exactly as unfinished as it was a year ago, and the contradiction is starting to show.
The per-token price is the easy line. Everything else, the retries, the context overhead, the agentic tool calls, the egress, the GPU reservation underneath the API, is where the actual bill comes from.
