An architect is designing a multi-agent system where a Claude orchestrator decomposes complex research tasks into subtasks delegated to specialized Claude subagents. If any subagent fails, the orchestrator must retry with a modified approach rather than propagating the failure. Where should retry logic be implemented?

In a multi-agent system, strategic retry and failure handling belongs in the orchestrator, not individual subagents. The orchestrator has the full task context needed to make intelligent decisions: it knows which subtask failed, what the overall research goal is, and can choose to retry with a different prompt, substitute an alternative subagent, or decompose the failed subtask differently. Option A (subagent self-retry) creates local retry loops without task context — a web search agent retrying the same failed search query gains nothing without the orchestrator's context about why the search was needed. Option C (API-layer backoff) handles transient infrastructure failures (5xx errors) but cannot handle semantic failures (wrong output, insufficient information) that require prompt-level correction. Option B creates conflicting retry logic and can cause retry storms when both layers retry simultaneously.