Claude Certified Architect – Foundations: study guide

The Claude Certified Architect – Foundations exam tests whether you can design agent systems with Claude, and every one of its 60 questions is set inside a production scenario. The guide lists 6 scenarios and draws 4 at random for each sitting, so you prepare for all 6 but answer questions from only 4. The largest domain is agentic architecture and orchestration, at 27% of the marks.

It takes 120 minutes and costs $125. The pass is a scaled 720 out of 1,000. Each question says how many answers to pick. Only people at organisations in the Claude Partner Network can register. Our comparison of the 4 Claude certifications covers who each one is for; this guide is about passing this one.

The 6 scenarios

The marks split across 5 domains: agentic architecture 27%, Claude Code configuration 20%, prompt engineering and structured output 20%, tool design and MCP 18%, context management and reliability 15%. Each scenario leans on 2 or 3 of them, so one missing scenario does not leave a whole domain out.

The answers the questions reward

Read the task statements in the guide and the same instincts come up again and again. If you hold these, many questions answer themselves.

• Let the API say when to stop. An agent loop goes on while stop_reason is tool_use and ends at end_turn. The guide names the wrong answers outright: searching the reply for words like "done", treating any text as the end, or making an iteration cap the main way out. Our sample exam opens with exactly this.
• An error is not an empty result. A tool that times out should say so, with isError and enough detail for the agent to decide whether to retry. An empty list means the search worked and found nothing.
• The coordinator chooses. A coordinator looks at the question and invokes only the subagents it needs, rather than running every question through the full pipeline.
• A fresh reviewer. The session that wrote the code tends to approve it. A separate instance given only the diff and the criteria catches more.
• Retries fix format, not missing facts. Sending the validation error back helps when the model got the structure wrong. It cannot help when the value is in a document you never sent.

Where the guide and today's Claude Code disagree

The guide is version 1.0 from July 2026, and Claude Code and the API have changed since. The exam marks the guide's answer, so learn that first and today's behaviour second. These are the gaps we found on 24 October 2026:

One more is worth knowing for the CI scenario. The guide treats CLAUDE.md as how a CI run gets project context, which is still true for plain claude -p. The headless docs now suggest --bare for scripted calls, though, and --bare skips CLAUDE.md entirely.

How to study

Spend the most time where the marks are. For agent loops, subagents and MCP, the tools and agents section of our AI engineering course builds them step by step. The Claude Code domain is about configuration rather than prompting. The project instructions, subagents and headless pages cover the files and flags the scenarios use. For structured output, know the difference between a schema the model is asked to follow and one that is enforced.

Then work through the scenarios one at a time, since that is how the questions arrive. The guide also lists 4 preparation exercises, starting with a multi-tool agent that escalates to a person; build them rather than reading about them. The free Architect – Foundations practice questions are written to the guide's task statements, one scenario per question. If your next step is the senior exam, the Professional guide covers the governance and stakeholder work this one leaves out.