Common workflow patterns for AI agents

Three core workflow patterns for AI agents in production are covered: sequential (for dependent tasks), parallel (for independent concurrent tasks), and evaluator-optimizer (for iterative quality refinement). Each pattern has distinct tradeoffs around latency, token cost, and reliability. The guidance recommends starting with the simplest approach—a single agent call—and only adding workflow complexity when measurable quality or performance gaps justify it. Patterns can be nested and combined as requirements grow, but complexity should always match actual needs.