Match the minimum complexity
Adding agents is distributed-systems complexity, not maturity. Multi-agent setups can burn ~15× the tokens of a single agent. No framework removes the need for authorization, audit and state layers.
Refuted 0–3. The blanket preference doesn't hold — it's a conditional choice, superseded by the minimum-complexity principle above.
Sequential · hierarchical/manager · parallel fan-out · adversarial/debate · reflection. LangGraph's explicit graph model gives the strongest control over applying them.
Interrupt-based (pause for approval before irreversible actions) and async-review (log for later audit). Treated as mandatory for financial, legal, safety or regulatory workflows.
Both fixed-list framings refuted 0–3. The building blocks are real — ReAct, Reflection, Plan-and-Execute, Tool Use, Memory, HITL — but there is no single canonical numbered list.