Core concept: Multi-agent systems work when each subagent keeps its own context and every failure stays visible. The anti-patterns are the choices that break isolation or hide errors from the coordinator.

Why A and B are correct:

A hides a real failure. Anthropic's tool guidance says error responses should "clearly communicate specific and actionable improvements, rather than opaque error codes". An empty result set reads as "no data exists", so the coordinator states a confident wrong conclusion.
B destroys context isolation. In the Agent SDK each subagent "runs in its own conversation, which starts fresh", and does not receive "the parent's conversation history or tool results". Shared global state refills every context window and couples the agents.
Anthropic's research system instead uses "an orchestrator-worker pattern, where a lead agent coordinates the process while delegating to specialized subagents that operate in parallel".

Why the others are wrong:

C: This is the recommended way to get parallelism. The lead agent "spins up 3-5 subagents in parallel rather than serially", which cut research time "by up to 90%".
D: Structured claim-source mapping is a best practice. It is how attribution survives the hand-off back to the coordinator.

Remember: Anti-patterns hide failures and share context. Good designs isolate context and surface every error.

Question 192