CCAR-F Domain 3: Multi-Agent & Orchestration Patterns
Designing modular AI architectures: Prompt Chaining, Routing, Parallelization, Orchestrator-Worker, and Evaluator-Optimizer loops.
1. The 5 Core LLM Architectural Patterns
1. Prompt Chaining
Decomposes a complex task into a linear sequence of steps where the output of step N becomes the input to step N+1. Ensures high precision at each stage.
2. Routing
Classifies input queries and directs them to specialized prompts or specialized models (e.g. routing simple math to Haiku, coding to Sonnet).
3. Parallelization
Runs independent sub-tasks concurrently (e.g. section-by-section document analysis, voting/consensus checks) to reduce total elapsed clock time.
4. Orchestrator-Worker
A central orchestrator LLM dynamically breaks a complex goal into tasks, delegates them to specialized worker sub-agents, and synthesizes the final result.
2. The Evaluator-Optimizer Loop
In high-precision tasks (such as code generation or complex translation), one LLM generates a draft solution, while a second LLM (the Evaluator) provides structured critique against a rubric. The first LLM refines its output in iterative loops until quality thresholds are satisfied.
3. When to Choose Simple Chains vs. Autonomous Agents
Architectural Rule of Thumb
Always default to deterministic workflows (Chaining / Routing) whenever the task steps are predictable. Reserve autonomous agentic loops for open-ended problem spaces where the sequence of tool calls cannot be hardcoded in advance.