← ClaudeAtlas

nw-design-patternslisted

7 agentic design patterns with decision tree for choosing the right pattern for each agent type
nWave-ai/nWave · ★ 541 · Web & Frontend · score 84
Install: claude install-skill nWave-ai/nWave
# Agentic Design Patterns ## Pattern Decision Tree ``` Is the agent doing a single focused task? YES -> Does it need self-evaluation? YES -> Reflection NO -> ReAct (default for most agents) NO -> Is it coordinating multiple agents? YES -> Are tasks independent? YES -> Parallel Orchestration NO -> Are tasks sequential with dependencies? YES -> Sequential Orchestration NO -> Hierarchical (supervisor + workers) NO -> Is it routing to one of several specialists? YES -> Router NO -> Does it need structured task decomposition? YES -> Planning NO -> ReAct (default) ``` ## 1. ReAct (Reason + Act) General-purpose agents needing tool calling and iterative problem-solving. **Loop**: Reason -> Select/execute action -> Observe result -> Repeat until done. **When**: Default pattern. Most specialist agents. **Examples**: software-crafter, researcher, troubleshooter. ## 2. Reflection Agent must evaluate and iteratively improve its own output. **Loop**: Generate -> Review against criteria -> Identify gaps -> Refine -> Validate threshold met. **When**: Quality-critical outputs where first-draft insufficient (code review, architecture review, agent validation). **Examples**: agent-builder-reviewer, solution-architect-reviewer, software-crafter-reviewer. ## 3. Router Request classified and delegated to exactly one specialist. **Loop**: Analyze request -> Classify -> Select specialist -> Delegate. **When**