CCAR-F is the architect's credential in Anthropic's certification program. It is written for people who design production systems on Claude rather than people who write a single integration, and the guide says so: a solution architect with about six months of hands-on work across the Claude API, the Agent SDK, Claude Code and MCP.
The exam is built out of scenarios. The guide publishes six production contexts, a support agent on MCP tools, a multi-agent research pipeline, Claude Code in a CI/CD pipeline, and three more, and a sitting draws four of them. Questions sit inside a context and ask what you would do about a number the scenario just gave you. A support agent resolving 55% of contacts against an 80% target is not a definition question.
The weighting is the thing to read first. Agentic Architecture and Orchestration is the largest domain at 27%, and its seven task statements run from the agentic loop and stop_reason handling through coordinator and subagent patterns to session forking. Claude Code Configuration and Workflows and Prompt Engineering and Structured Output take 20% each, Tool Design and MCP Integration 18%, and Context Management and Reliability 15%.
Most items give you a production symptom and four fixes that would all help. One addresses the cause the scenario named. The recurring judgment is when a prompt instruction is enough and when a business rule needs programmatic enforcement, which the guide's own sample questions turn on more than once.
The platform moves fast, which is what older material gets wrong. These tests are written against the exam guide dated July 2026 and the Claude platform as it stands, including the Agent SDK, hooks, MCP, the Claude Code configuration hierarchy and the Message Batches API.