Skip to content

Part 7 — Summary

Part 7 closed the lifecycle. What was established:

  • Environments in AI involve more than application copies: models, prompts, RAG indexes, tools, policies and data, each with its own cycle.
  • AI-specific CI/CD requires an eval gate on all artifacts that affect behavior: models, prompts, tools, MCP servers, indexes, agents, policies.
  • Versioning covers everything that can affect behavior; a behavior change without a schema change is still a breaking change.
  • Rollout combines canary, shadow, feature flag, blue/green and progressive rollout; with reversal criteria written beforehand.
  • Deprecation is engineering, with deterministic blocking, a window and audit — not a flag in metadata.
  • Breaking-change communication reaches internal users, consuming agents, MCP servers and A2A peers.
  • Fine-tuning is a technique useful for format, style and specific behavior, not for factual knowledge or business rules.
  • ML at scale (distributed training, FSDP/ZeRO) is a strategic topic; most teams consume ready-made models and do not need this infrastructure.

Part 8 reorganizes what comes after this book from a practical standpoint: how the planned examples connect to the chapters.