The Five Reference Architectures for the CCAR-P Exam

Shapes the industry already paid to learn

Most partner problems map to one of a handful of reference architectures already proven in the Claude ecosystem. The Claude Certified Architect - Professional (CCAR-P) exam treats recognising these five as an understand-level skill because they are the vocabulary of solution design: knowing them well enough to generalise from lets you take the shape that fits, adapt it to the workload in front of you, and recognise when a problem draws on more than one. A reference architecture is a documented pattern for wiring an LLM application to solve a recurring class of problem, and the point is to understand why each works, not to memorise a diagram.

The five are: the agent, an open-ended loop where Claude decides its own next step until a stop condition; RAG, where retrieval augments a call with relevant chunks from a stable corpus; the document-processing pipeline with an evaluator-optimizer pass, which processes items and then checks output against a rubric and refines; routing, where an initial classification step sends work down a specialised path; and the coding agent, which pairs agentic exploration with deterministic edit, test, and review steps.

The five shapes and what each is for

The agent suits open-ended tasks where the sequence of steps cannot be predetermined, so Claude plans and executes toward a goal, deciding what happens next until a stop condition fires. RAG suits questions answerable from a stable body of reference material, documentation, policy text, product manuals, where retrieving the relevant chunks and handing them to the model produces grounded answers. The document-processing pipeline with an evaluator-optimizer pass suits high-volume processing where output quality must be checked: each item is processed, then an evaluator scores it against a rubric and triggers a refinement pass when it fails.

Routing suits work that splits into distinct intents, an initial classification step sends each request down a specialised path, as in customer-service ticket triage. The coding agent suits software tasks: it explores a codebase agentically but pairs that exploration with deterministic edit, test, and review steps so the open-ended reasoning is bounded by checkable operations. Each shape encodes a hard-won answer to a recurring problem, which is why starting from the right one saves the cost of rediscovering it.

Adapt the shape, do not copy the blueprint

The central discipline is treating a reference architecture as a shape to adapt, not a template to implement as drawn. Every partner workload is unique, so the goal is to understand the pattern well enough to reshape it to the problem, adjusting where the standard shape does not fit. A reference architecture implemented rigidly, without adapting it to the actual data and query pattern, imports assumptions that may not hold. The reference is a starting point that carries the accumulated learning of prior projects, not a finished answer to the current one.

Habit is the enemy of fit

The second discipline is not picking a reference architecture out of habit. Reusing RAG because it worked on a prior project, without checking whether this problem's data and query pattern match, is how retrieval gets applied to live state, the single most common reference-architecture mistake. RAG earns its place on stable corpora; it breaks on data a live system owns and updates. The check is to ask, each time, whether the properties that make a given shape appropriate actually hold for this workload, rather than reaching for the shape that is most familiar.

What the CCAR-P exam trips candidates on

The exam tests two traps. The first is assuming a reference architecture is a rigid template to implement as drawn, rather than a shape to adapt to the workload. A scenario will show a standard pattern applied literally to a problem it does not quite fit, and the credited answer adapts the shape rather than forcing the problem into it.

The second is picking a reference architecture out of habit, such as reusing RAG from a prior project, without checking whether this problem's data and query pattern match it. The reliable reading interrogates the fit: does the data have the property, stability for RAG, distinct intents for routing, that the shape assumes? If not, the familiar shape is the wrong one.

Common misreadings to avoid

How this shows up on the exam

Domain 1 questions on this knowledge point ask you to name the reference architecture that fits a described problem, or to critique a chosen one. The reliable reading recognises the five shapes, agent, RAG, document pipeline with evaluator-optimizer, routing, coding agent, treats each as adaptable, and checks whether the workload's data and query pattern match the shape's assumptions rather than defaulting to a familiar pattern.

These shapes build on the four-stage architecture shape and on live state vs static knowledge, which decides whether RAG or a tool call belongs. They feed forward into composing two reference architectures deliberately and overlap with the workflow sub-patterns that routing and evaluator-optimizer draw on.