Skip to content
Hands-on AI Playbook

Workflow Architecture Patterns

Every AI workflow falls somewhere on an autonomy spectrum. These seven patterns — drawn from Anthropic’s research — provide a common vocabulary for describing how AI systems are structured, from simple tool-augmented models to fully autonomous agents.

The right pattern depends on what your workflow actually needs, not on how sophisticated you want it to be. Start simple, upgrade when you hit limits.

The Autonomy Spectrum

Section titled “The Autonomy Spectrum”

The patterns are organized into three tiers of increasing autonomy:

Foundation

Section titled “Foundation”
PatternDescription
Augmented LLMAn LLM enhanced with retrieval, tools, and memory — the building block for all other patterns

Structured Workflows

Section titled “Structured Workflows”
PatternDescription
Prompt ChainingBreak a task into sequential steps, with validation gates between each step
RoutingClassify input and direct it to a specialized follow-up process
ParallelizationRun subtasks simultaneously and aggregate the results
Orchestrator-WorkersA central orchestrator dynamically breaks down tasks and delegates to specialized workers
Evaluator-OptimizerGenerate output, evaluate it against criteria, and refine iteratively until it meets quality standards

Autonomous

Section titled “Autonomous”
PatternDescription
Autonomous AgentsAn LLM with tools, memory, and planning that independently executes multi-step tasks through a think-act-observe loop

Pattern Selection Framework

Section titled “Pattern Selection Framework”

Use these three questions to identify which pattern your workflow needs:

1. Is the task predictable or open-ended?

  • Predictable (you can define the steps in advance) → Use a structured workflow pattern
  • Open-ended (steps depend on what the AI discovers) → Consider an autonomous agent

2. How many steps are involved?

  • Single step → Augmented LLM
  • Sequential steps → Prompt Chaining
  • Branching paths → Routing
  • Independent parallel steps → Parallelization
  • Dynamic subtasks → Orchestrator-Workers

3. Does the output need iterative refinement?

  • Yes, with clear quality criteria → Evaluator-Optimizer
  • Yes, with open-ended exploration → Autonomous Agent

Decision Flow

Section titled “Decision Flow”
Start here: Can you define all the steps in advance?
│
├── YES → How many steps?
│   ├── One step → Augmented LLM
│   ├── Sequential steps with validation → Prompt Chaining
│   ├── Input determines the path → Routing
│   ├── Independent steps that can run simultaneously → Parallelization
│   └── Steps need dynamic decomposition → Orchestrator-Workers
│
├── PARTIALLY → Does output need iterative refinement?
│   └── YES → Evaluator-Optimizer
│
└── NO → Steps are unknown until runtime
    └── Autonomous Agent

How These Patterns Relate to the Framework

Section titled “How These Patterns Relate to the Framework”

The Design phase of the Business-First AI Framework first assesses the workflow’s autonomy level (Deterministic → Guided → Autonomous), then recommends an orchestration mechanism (Prompt → Skill-Powered Prompt → Agent). These seven architecture patterns provide the detailed implementation blueprints within that spectrum:

Orchestration MechanismArchitecture Patterns
PromptAugmented LLM
Skill-Powered PromptPrompt Chaining, Routing
AgentParallelization, Orchestrator-Workers, Evaluator-Optimizer, Autonomous Agents

The AI Workflow Design Matrix adds a second dimension — human involvement (Augmented vs. Automated) — to the autonomy spectrum. Architecture patterns map to the matrix’s autonomy axis: Deterministic workflows typically use Augmented LLM or Prompt Chaining, Guided workflows use Routing or Evaluator-Optimizer, and Autonomous workflows use Orchestrator-Workers or Autonomous Agents.

Credit

Section titled “Credit”

Based on Building Effective Agents by Anthropic.

Section titled “Related”