The Claude Certified Architect Foundations exam tests how you reason about agentic systems in production - agent loops that don't terminate, MCP servers that fail under multi-client load, prompts that break on the next model upgrade, Claude Code configurations that drift across an engineering org. The official documentation tells you what each surface does. This book tells you which surface to reach for, where it breaks, and what survives the next release.
Eighteen chapters. Five domains, weighted to the published blueprint. One opinionated argument throughout: the senior engineer's job is the durable architectural reasoning around a moving-target dependency, and the cert is testing exactly that.
What this book covers:
- The agent loop as an architectural commitment - when it earns its complexity, when a deterministic pipeline is cheaper, and the four termination conditions every production loop needs.
- Tool design as API design the model will misuse - descriptions as prompts, parameter names that carry contracts, return shapes that compose with the next decision.
- MCP from first principles - primitives, lifecycle, security, the seven disciplines that separate spec-conformant servers from production-ready ones.
- Prompt engineering as interface design - system prompts under source control, validation strategies that fit the operation, the patterns that survive model upgrades.
- Claude Code as infrastructure - the configuration hierarchy, the composition model for skills/subagents/hooks/plugins, headless invocation in CI with the bounds that prevent cost incidents.
- Context as a budget - compaction, retrieval, prompt caching, and the failure modes of long context that present as model regressions.
- Evals you'll actually run - defined dataset, defined task, defined scoring, defined aggregation, with the cadences that catch what offline misses.
The back matter includes the full exam blueprint cross-referenced to chapter and section, sixty practice questions with reasoning for each correct answer and the trap behind each distractor, a twelve-week study plan with parallel read/build/evaluate/review lanes, and three end-to-end reference architectures with decision logs and named failure modes.
Real architectural decisions. Real failure modes. The reasoning that holds across model upgrades, blueprint revisions, and the next surface Anthropic ships.
For senior engineers, staff+ ICs, and architects sitting the CCA-F exam - and for the readers who want the architecture book the cert apparatus is built around.
Not endorsed by, affiliated with, or sponsored by Anthropic. Practice questions are calibrated to the publicly documented blueprint and are not leaked exam items.
