Coze Multi-Agent Orchestration

Choosing the wrong multi-agent workflow wastes the execution cost of every downstream node, so success was defined as accurately identifying applicable boundaries and distinguishing easily-confused scenarios. Defined metadata for task type, applicable scenarios, agent roles, I/O, process nodes and constraints, and built a 30-case evaluation set including confusable items. Bad-case attribution pointed at insufficient model knowledge, so the engineering layer gained 5-shot examples with positive/negative cases and reasoning explanations, using an ICE-structured prompt with schema-structured output to avoid hallucination.