Anthropic launched a Claude Certified Architect exam. Here's what actually matters.

Anthropic quietly rolled out a Claude Certified Architect Foundations cert. It’s free for the first 5,000 partner network members.

I dug into the official exam guide. Here’s what stood out.

The exam weights tell you where Anthropic thinks the skill gap is:

• Agent architecture and orchestration: 27%

• Claude Code configuration: 20%

• Prompt engineering and structured output: 20%

• Tool design and MCP: 18%

• Context management and reliability: 15%

Not prompt tricks. Not API calls. Orchestration. How agents talk to each other, pass context, and fail safely.

The sharpest idea in the whole guide:

Hooks are deterministic. Prompts are probabilistic.

If your rule is “don’t refund over $500,” a prompt gives you ~95% compliance. A hook gives you 100%. Use hooks for policy. Use prompts for style. Most teams get this backwards.

Three other things worth stealing:

1. Tool descriptions are how the model picks tools. Vague descriptions = wrong tool 12% of the time. Rewrite them like API docs.

2. Subagents don’t inherit context. You have to pass it explicitly in the prompt. Forget this and your multi-agent system silently breaks.

3. stop_reason is your agent loop’s only reliable stop signal. Not text parsing. Not max iterations. Just stop_reason.

My take:

Certs usually lag real practice by 2 years. This one doesn’t. It’s testing harness engineering, which is exactly what I’ve been doing at Popp and on my own builds.

Worth the $99 if you’re shipping with Claude. Worth reading the free guide either way.

Link Guide