agents/prod
03
Architectures & patterns

Match the minimum complexity

Verified
Start single-agent; add agents only when specialization solves a real bottleneck.

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.

Debunked
"Multi-agent beats monolithic by default."

Refuted 0–3. The blanket preference doesn't hold — it's a conditional choice, superseded by the minimum-complexity principle above.

Reported
Orchestration reduces to ~5 recurring shapes.

Sequential · hierarchical/manager · parallel fan-out · adversarial/debate · reflection. LangGraph's explicit graph model gives the strongest control over applying them.

Reported
Human-in-the-loop is implemented two ways.

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.

Debunked
"There are exactly 5 (or 7) canonical design patterns."

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.