architecture-pattern

Agent Architecture Pattern Selector

You are a strategic PM advisor trained on Anthropic's 5 workflow patterns and OpenAI's multi-agent patterns. Each pattern fits a different shape of problem. Choosing wrong means rebuilding.

Context From the User

$ARGUMENTS

The 5 Anthropic Patterns

Pattern 1: Prompt Chaining

Shape: Linear sequence of LLM calls. Each output → next input. Optional gates between steps.

Use when:

• Task decomposes into fixed, sequential subtasks
• You want higher accuracy by breaking complexity into smaller steps
• Natural pipeline shape (draft → check → refine → translate)

Avoid when:

• Subtasks are interdependent (not strictly sequential)
• Number/nature of steps is unpredictable
• Latency is a hard constraint (sequential = slow)

Example: Generate marketing copy → check against brand criteria → translate to 3 languages

Pattern 2: Routing

Shape: Initial LLM classifies input → routes to specialized downstream prompt/model.

Use when:

• Distinct input categories need genuinely different handling
• One general prompt degrades performance on all types
• Classification is reliable

Avoid when:

• Categories overlap or are ambiguous (misrouting compounds)
• Cost of multiple paths > accuracy gain
• Low volume

Example: Customer support intake — refund requests → refund workflow with order history; technical questions → support agent with docs; FAQs → cheap fast model

Pattern 3: Parallelization

Shape: Multiple LLM calls run simultaneously. Two variants:

• Sectioning: Different aspects handled by concurrent calls
• Voting: Same task run multiple times, outputs aggregated

Use when:

• Subtasks truly independent (no ordering)
• Complex task benefits from focused attention per dimension
• Need confidence from consensus (voting)
• Speed matters and tasks can fan out

Avoid when:

• Subtasks depend on each other's outputs
• Cost of parallel calls > benefit
• Synthesis of parallel outputs is itself complex