Anthropic's Claude Certified Architect – Foundations exam is scenario-based: 60 questions, 120 minutes, 720/1,000 to pass. It's not an "AI literacy" badge — it tests production judgment. Here's the domain breakdown and the specific traps in each.
Agentic Architecture & Orchestration — 27%
The biggest domain. The agentic loop is simple in principle: stop_reason: "tool_use" means execute the tools and loop back; stop_reason: "end_turn" means stop. The trap is anything else — checking for "done" in the assistant's text, or capping at N iterations as your primary control. A hard cap is a safety net, not the signal.Claude Code Configuration & Workflows — 20%
Lots of CLAUDE.md and CI/CD. A favorite gotcha: a Claude Code job in CI hangs forever because it was launched without -p. Without the print flag it starts an interactive session and waits on stdin that never comes. -p runs headless: process, print to stdout, exit.Prompt Engineering & Structured Output — 20%
Few-shot patterns, JSON schemas, output validation. The exam assumes you can prompt; it tests whether you can make output reliable.Tool Design & MCP Integration — 18%
Tool descriptions, MCP servers, error handling. Key principle: when a tool fails, return a structured error the model can read and recover from — don't throw and kill the loop, and don't return an empty string it'll silently misread.Context Management & Reliability — 15%
Long sessions, escalation, provenance, and a subtle one: the same Claude instance that generated code will rationalize its own decisions when asked to review them — an independent instance catches more.
How to prep: figure out your weak domain first. I put together a free, no-signup diagnostic that scores you across all five in ~10 minutes: https://www.claudecertifiedarchitects.com/
Independent study resource — not affiliated with or endorsed by Anthropic. Verify current exam details against the official guide before you sit it.
Top comments (0)