Skip to content
back to the archive page
#AI4SDLC

Claude Certified Architect: The Exam as a Map of Agent Engineering (Category AI4SDLC)

#AI4SDLC #AI #Agents #Architecture #Engineering #DevTools

I watched a 20-minute talk by Frank Coyle at AI Engineer World's Fair 2026 about the Claude Certified Architect — Foundations exam. It is more useful to read it not as a certificate but as a map of the decisions without which an agent demo never becomes a production system. Incidentally, I previously covered Frank's talk on ontologies from the same conference.

CCAR-F is Anthropic's first technical certification for architects building applications with Claude. It consists of 60 single- or multiple-answer questions over 120 minutes; the passing scaled score is 720 out of 1,000. Four of six scenarios are selected at random: support, code generation, multi-agent research, developer tools, Claude Code in CI/CD, and structured extraction. The exam tests not a product menu but architectural choices made with constraints and failure modes in mind.

The exam topics reveal its priorities clearly:

  • Agent architecture and orchestration — 27%: loops, stop_reason, subagents, and hooks;
  • Tools and MCP — 18%: schemas, errors, and tool allocation;
  • Claude Code — 20%: CLAUDE.md, skills, plan mode, and CI/CD;
  • Prompts and structured output — 20%: few-shot examples, JSON Schema, and validation;
  • Context and reliability — 15%: compaction, data provenance, escalation, and human review.

Coyle explains the exam map through antipatterns. Do not mistake a model response for an action: the LLM produces a tool call, while code checks stop_reason, executes the call, and returns the result to the loop. Do not give one agent every tool if a narrow subagent is enough. Do not dump its complete history into the main context: return conclusions, facts, and links, and compact long sessions. Coyle spends considerable time on loop engineering, which Max Smirnov and I discussed in Research Insights Made Simple #19. There we talked about a repeatable loop and the reviewer's authority to stop it. CCA adds a map of the engineering competencies around that loop: tools, context, boundaries, verification, and recovery.

In short, the value of this exam lies not in the badge but in its preparation framework, which can be used as a curriculum and an architecture-review checklist: where does execution stop, who executes a tool call, what can the subagent see, is provenance preserved, and when is a decision handed to a person? The map is tied to Claude, and passing the test does not prove production operating experience. The conditions described in the talk are already out of date: as of August 12, 2026, the exam price had risen to $125, but its questions remain useful even for people who do not plan to take it.

#AI #AI4SDLC #Agents #Architecture #Engineering #DevTools