Augmented LLM
A single model call enhanced with retrieval, tools, and memory. The basic building block of Anthropic's architectural taxonomy, and for most business problems the whole answer.
Exam context: The default tier. Escalation to workflows or agents must be justified by a requirement the augmented LLM cannot meet.
See also: 1.3 Selecting Architectural Patterns
Workflow
A system where LLMs and tools are orchestrated through predefined code paths. The developer decides the steps; the model executes within them. Predictable, testable, and auditable.
Exam context: The correct choice when steps are known in advance. Named patterns: prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer.
See also: 1.3 Selecting Architectural Patterns
Agent
A system where the model dynamically directs its own processes and tool usage, looping on environment feedback until the goal is met. Flexible, at the price of cost, latency, a wider error surface, and a harder audit story.
Exam context: Earns its cost only when the path genuinely cannot be scripted. The impressive-sounding agent for a fixed pipeline is the exam's favourite over-engineering distractor.
See also: 1.3 Selecting Architectural Patterns
The Fit Test
The design question of whether a task belongs with a model at all. Language understanding, synthesis, and judgement over unstructured input justify Claude; fully rule-specified tasks belong in deterministic code.
Exam context: Recognising where NOT to use a model is scored as highly as recognising where to use one.
See also: 1.1 Translating Business Problems
Success Criterion
A metric with a baseline and a target, defined before the build, that makes a solution's value measurable and falsifiable. Without one, a solution is a demo.
Exam context: The professional sequence is problem → constraints → success criterion → solution shape. Options that jump to a build with no baseline are planted wrong answers.
See also: 1.1 Translating Business Problems
Six-Stage Pipeline
The canonical production structure: input, context assembly, processing, output, delivery, feedback loops. Every production Claude system has all six, whether or not they were drawn.
Exam context: Gap-analysis questions describe an architecture with a stage missing. The two most commonly missing are output validation and the feedback loop.
See also: 1.2 End-to-End Architectures
Output Validation
Schema checks plus business-rule checks applied to model output before downstream systems consume it. Model output is untrusted input until validated; schema validity is not business validity.
Exam context: The control that stops bad model output corrupting downstream state. Prompt instructions are guidance, not a guarantee, and never substitute for it.
See also: 1.2 End-to-End Architectures
Feedback Loop
The pipeline stage that captures quality signals — user feedback, sampled evaluations, monitoring — and routes them into prompt, retrieval, and design iteration. The stage that separates production systems from demos.
Exam context: "Quality is inconsistent and nobody can say why" is the missing-feedback-loop signature. Upgrading the model is the planted distractor.
See also: 1.2 End-to-End Architectures
Blast Radius
The downstream consequence of a wrong model output. Determines how much validation and human review an architecture needs, and drives the delivery choice (auto-send vs draft-for-review).
Exam context: High blast radius (money, compliance, customer trust) demands prevention before delivery, not detection after it.
See also: 1.2 End-to-End Architectures
Orchestrator-Worker Architecture
A multi-agent shape where a lead agent decomposes a goal, spawns specialised subagents that work in parallel with separate context windows, and synthesises their results. Communication flows through the coordinator.
Exam context: Delegation quality determines system quality. Duplication plus gaps in the output traces to the coordinator's decomposition, not to subagent capability.
See also: 1.4 Multi-Agent Systems
Context Isolation
The design fact that subagents inherit nothing and share nothing: no coordinator conversation, no peer visibility, no shared memory. Every piece of context a subagent needs must be passed explicitly in its brief.
Exam context: Distractors assume ambient shared knowledge between agents. Designs that rely on it are broken by construction.
See also: 1.4 Multi-Agent Systems
Token Multiplier
The measured cost of agentic architectures: agents run at roughly 4× chat token usage, multi-agent systems at roughly 15×. The entry fee a task's business value must clear.
Exam context: A multi-agent proposal with no value case is an incomplete answer. The exam tests restraint.
See also: 1.4 Multi-Agent Systems
Attention Dilution
The failure mode where one model pass over too much material produces inconsistent depth: thorough on some units, blind on others, contradictory on identical patterns. Fixed structurally with per-unit local passes plus a cross-unit integration pass.
Exam context: The flagship decomposition failure. A larger context window is the planted distractor — it solves fit, not attention.
See also: 1.5 Decomposition Techniques
Decomposition Axes
The four ways to split a complex task: by stage (sequential phases), by section (independent units), by perspective (different lenses on the same input), by specialism (different tools, prompts, or models per subtask).
Exam context: Choosing the axis is the design decision; the workflow pattern follows from it.
See also: 1.5 Decomposition Techniques
Acceptance Gate
A programmatic check between pipeline stages that stops bad intermediate output propagating. Checkable in code without a model call, with a defined rejection path.
Exam context: Errors compound across stages if unchecked. Gate-free pipelines offered as complete are under-engineered.
See also: 1.5 Decomposition Techniques
Business Value Pillars
The five ways a Claude solution creates value, per the CCAR-P guide: efficiency, transformation, productivity, cost, and performance SLAs. Every major design decision should name the pillar it serves and how that will be measured.
Exam context: Which-pillar questions plant adjacent pillars as distractors. Discriminator: automating an existing process is efficiency; a newly possible capability is transformation.
See also: 1.6 Business Value Pillars
Prompt Caching
Reusing a static prompt prefix (system prompt, policy documents, tools) across requests, cutting cache-read cost by roughly 90% and reducing time to first token. Static content goes first, dynamic content after.
Exam context: The double-pillar lever: serves cost and the latency SLA on the request path. Distractors swap it with batch processing.
See also: 1.6 Business Value Pillars
Batch Processing
Submitting large volumes of requests for asynchronous processing via the Message Batches API at roughly half the cost. Serves the cost pillar for work that tolerates deferred completion.
Exam context: The off-request-path lever. Offering it for an interactive use case with a latency SLA is the classic swapped-lever distractor.
See also: 1.6 Business Value Pillars
SLA Engineering
Translating contractual latency, availability, and quality targets into architecture: p95 budgets decomposed into per-stage budgets, availability targets forcing designed fallbacks, quality floors requiring evaluation machinery.
Exam context: An SLA is engineered, not hoped for. A chain whose stages cannot sum inside the latency target must change shape.
See also: 1.6 Business Value Pillars