Claude Certified Architect

Richard M. Thompson, who founded AR Labs, holds Claude Certified Architect – Foundations, Anthropic's certification for people who design and build systems with Claude: agentic applications, tool integrations through the Model Context Protocol (MCP), Claude Code workflows, structured output and context management. Richard was awarded the credential in September 2026; it is valid until September 2027. While preparing, we built a study tracker for the exam as a small web application, which anyone studying for the same exam can use.

The exam

The exam is scenario-based and multiple-choice. Its published guide divides the material into five weighted domains and 30 task statements, each task statement listing the knowledge and skills it expects.

Domain	Share of the exam
Agentic architecture and orchestration	27%
Tool design and MCP integration	18%
Claude Code configuration and workflows	20%
Prompt engineering and structured output	20%
Context management and reliability	15%

How we prepared

The preparation started from practice rather than from courses. We already ran Claude Code, Anthropic's coding agent (an AI agent that writes code), daily with layered instruction files, skills (packaged instructions an agent loads only when a task needs them), MCP servers, sub-agent delegation and hooks (scripts run automatically around an agent's action that can block it), which covers much of the first three domains. Study time went to the gaps between that experience and the exam's level, mostly controls that Claude Code hides from its users: how stop_reason drives an agentic loop, the tool_choice modes, the Message Batches API, structured error responses for tools, prompt caching, and running Claude Code headless in CI pipelines.

We also mapped all 30 task statements onto the build phases of Salesbot, our browser voice agent, so that building it doubled as exam practice. About 22 of the 30 statements are exercised by that build. The remaining eight are specific to Claude's own runtime (Agent SDK mechanics, MCP and the batch API) and got targeted exercises instead of bending the voice agent's design to cover them.

The study fed back into our engineering. Two task statements, on intercepting tool calls with hooks and on enforcing multi-step workflows with handoff patterns, name exactly where we moved enforcement in our coding pipeline, the route every code change takes: out of prompt text and into code that can refuse a step.

The study tracker

The tracker mirrors the exam guide's three-level structure: domain, then task statement, then individual knowledge and skill items.

flowchart LR
  D["5 domains"] --> T["30 task statements"]
  T --> K["Knowledge items"]
  T --> S["Skill items"]
  K --> P["Ticked progress"]
  S --> P

It has three views:

• Browse. Every domain and task statement, with each knowledge and skill item as a checkbox.
• Working set. The task statements a person has chosen to focus on now, gathered in one place.
• Dashboard. Progress across all five domains.

Progress is saved in the visitor's own browser, so there is no account and nothing is sent anywhere. The page is part of the site's Next.js codebase. The exam content is transcribed from the official guide into a typed data file, and saved progress is keyed to an identifier for each item.