Claude Certified Engineers logoCCE

CCAR-P · Study Guide

← Domain 1: Solution Design & Architecture

1.3 · Lesson 3 of 6

Select appropriate architectural patterns (workflow, agentic, augmented LLM)

What You Need to Know

Architectural patterns for Claude solutions are chosen from task structure, control needs, latency, and unpredictability — not from buzzwords. Three core patterns cover most designs: workflow (fixed gates), agentic (open tool plans under a bound), and augmented LLM (single generation with retrieval or light tools). Prefer the simplest pattern that meets quality and control requirements.

Workflow fits sequential steps with explicit gates and deterministic checks. Agentic fits open-ended tool use and adaptive plans when the path cannot be fully sequenced upfront. Augmented LLM fits bounded Q&A and light tooling around one call. Hybrid designs are allowed when stages have different control needs — for example, augmented retrieval inside a workflow commit gate.

Pattern cheat sheet

  • Workflow — fixed sequence, explicit gates, deterministic checks, strong audit trail
  • Agentic — adaptive tool plans, bounded toolkit, open exploration under policy
  • Augmented LLM — retrieve / light tools + one generation, bounded unpredictability
  • Hybrid — different patterns per stage when control needs diverge

Decision rules

  • Workflow fits fixed sequential steps with explicit gates and deterministic checks.
  • Agentic fits open-ended tool use and adaptive plans under a bounded toolkit.
  • Augmented LLM fits retrieval or light tools around a single generation call.
  • Prefer the simplest pattern that meets control and quality requirements.
  • Hybrid designs are allowed when stages have different control needs.

Why simplest-that-fits wins

Extra agency without need adds selection noise, audit gaps, and latency. Extra workflow rigidity on truly open research forces brittle graphs that break when the next step depends on findings. Architects match pattern to unpredictability and control: claims gates favor workflow; investigative support may favor agentic; FAQ-style policy answers often favor augmented LLM.

Review checklist

  • Are steps known and gated, or does the path depend on intermediate findings?
  • Which checks must stay deterministic in code?
  • Is a single augmented call enough, or is multi-step irreversible branching required?
  • Would a simpler pattern still meet audit and quality bars?
  • If hybrid, which stage uses which pattern and why?

Exam distractors push always-multi-agent, agentic-for-fixed-pipelines, or one augmented call for complex irreversible branching. The keyed move is pattern-to-structure match.

Exam application

Fixed gates + deterministic checks → workflow. Open exploration under a toolkit → agentic. Bounded Q&A → augmented LLM. Prefer simplest fit; hybrids OK by stage. Distractors: always multi-agent, agentic for fixed pipelines, augmented for complex irreversible branching.

Exam traps

  • Always multi-agent by default

    Multi-agent is justified by role isolation or open exploration needs — not by fashion. Fixed pipelines usually want workflow.

  • Agentic for fixed pipelines

    Open tool plans fight auditability when steps and checks are already known. Prefer workflow with explicit gates.

  • Augmented LLM for complex irreversible branching

    Single-call retrieval-plus-generate under-controls multi-step irreversible decisions. Use workflow or staged hybrids.

  • Rejecting hybrids when stages differ

    Hybrids are allowed: e.g. augmented Q&A inside a workflow gate, or a bounded agentic research step before a fixed commit gate.

Practice scenario

An insurer needs a claims intake process with fixed sequential gates: validate forms, run deterministic eligibility checks, request missing documents, then draft a summary for a human adjuster. Control and auditability matter more than open exploration. Which pattern should the architect select?

Choose one answer

Build exercise

Pick a pattern for three scenarios and write one-sentence rationales

30 minutes

What you'll learn

  • Score scenarios on control, latency, and unpredictability
  • Select workflow for fixed gated processes with deterministic checks
  • Select agentic for open-ended tool plans under a bound
  • Select augmented LLM for bounded Q&A and allow hybrids by stage

  1. Step 1

    Score three scenarios on control, latency, and unpredictability

    For each scenario, note whether steps are known upfront, whether checks are deterministic, how open-ended tool use must be, and latency/audit constraints.

    Why: Pattern choice follows task structure — not buzzwords. The exam keys on matching control needs.

    You should see: A scorecard with control / latency / unpredictability notes per scenario.

  2. Step 2

    Pick a pattern for a fixed gated claims-like process

    Choose workflow. Place deterministic checks in code gates. Use Claude only for language steps (draft summary, classify ambiguous notes) inside the sequence.

    Why: Fixed sequential gates with deterministic checks are the textbook workflow fit.

    You should see: One-sentence rationale: workflow because gates and checks are fixed and auditable.

  3. Step 3

    Pick a pattern for open-ended research with tools

    Choose agentic with a bounded toolkit when the next action depends on intermediate findings and the path cannot be fully sequenced in advance.

    Why: Agentic fits adaptive plans under explicit tool bounds — not unbounded swarms.

    You should see: One-sentence rationale naming bounded tools and why the path is open-ended.

  4. Step 4

    Pick a pattern for bounded policy Q&A and allow hybrids

    Choose augmented LLM for retrieve-then-answer. If a later irreversible action appears, wrap that stage in a workflow gate rather than forcing the whole product into one pattern.

    Why: Simplest pattern that meets needs wins; hybrids are OK when stages differ.

    You should see: Three one-sentence rationales (workflow / agentic / augmented) ready for design review.

Sources