Appendix B: Pattern Catalog

Pattern	Description	Use when	Main risk	Key Reference
Single-shot answer	One model call produces the final answer	Task is simple and self-contained	Cannot recover from missing information	(39)
Retrieve-then-answer (RAG)	Retrieve relevant facts, then answer	Knowledge is external and large	Retrieval misses relevant facts	(12)
ReAct loop	Reason, act, observe, repeat	Task requires multi-step interaction	Looping and context growth	(19)
Plan-then-execute	Produce a plan, then execute steps	Task has stable subtasks	Plan becomes stale as environment changes	(40)(41)
Reflect-verify	Check intermediate or final output before continuing	Correctness is critical and checkable	Verification becomes expensive or weak	(42)(43)
Human-in-loop	Pause for human approval or input	Actions are risky or ambiguous	Latency and user burden	(2)(44)
Summarize-continue	Compress history before continuing	Long tasks exceed context budget	Important details are lost in summary	(10)(11)
Agent-as-tool	Wrap an agent so it can be called like a tool	Specialization, isolation, or parallelism is needed	Cost, latency, and error compounding	(45)(46)