Skip to content

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

The patterns are organized into three tiers of increasing autonomy:

Foundation

Pattern Description
Augmented LLM An LLM enhanced with retrieval, tools, and memory — the building block for all other patterns

Structured Workflows

Pattern Description
Prompt Chaining Break a task into sequential steps, with validation gates between each step
Routing Classify input and direct it to a specialized follow-up process
Parallelization Run subtasks simultaneously and aggregate the results
Orchestrator-Workers A central orchestrator dynamically breaks down tasks and delegates to specialized workers
Evaluator-Optimizer Generate output, evaluate it against criteria, and refine iteratively until it meets quality standards

Autonomous

Pattern Description
Autonomous Agents An LLM with tools, memory, and planning that independently executes multi-step tasks through a think-act-observe loop

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

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

Start simple, upgrade when needed

Most workflows start as a single Augmented LLM or Prompt Chain. Only add complexity when you hit a concrete limitation — not because the problem seems complex.

How These Patterns Relate to the Framework

The Build > Design phase of the Business-First AI Framework uses an Execution Pattern Spectrum (Prompt → Skill-Powered Prompt → Single Agent → Multi-Agent) to classify workflows at a higher level. These seven architecture patterns provide the detailed implementation blueprints within that spectrum:

Framework Execution Pattern Architecture Patterns
Prompt Augmented LLM
Skill-Powered Prompt Prompt Chaining, Routing
Single Agent Parallelization, Orchestrator-Workers, Evaluator-Optimizer
Multi-Agent Orchestrator-Workers (multi-agent), Autonomous Agents

Credit

Based on Building Effective Agents by Anthropic.