CASE EVIDENCE / A RECORD
Lemony AI (lemony-ai) 公开的智能体工作流案例，来源为 公开代码库，复核于 2026-06-26T13:17:46Z。
原始记录：CascadeFlow: Claude 3 Haiku as Speculative Execution Model in Agent Cascading Runtime
Chinese Brief
Lemony AI (lemony-ai) 公开的智能体工作流案例，来源为 公开代码库，复核于 2026-06-26T13:17:46Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
任务
这是一个围绕智能体流程编排的真实任务，公开材料可以回溯到具体使用者和具体产物。 原始资料写作：CascadeFlow is an agent runtime intelligence layer (2,740 GitHub stars) that dynamically selects the optimal LLM for each query or tool call through speculative execution. It uses Claude 3 Haiku as the initial speculati…
公开材料提供公开代码、README 或项目配置，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：Per-query optimal model routing decisions; 40-85% cost reduction while retaining 96% quality of flagship models; sub-5ms routing overhead in agent execution loops
Claude 3 Haiku 在该案例中承担智能体流程编排相关的生成、分析、编排或实现角色。 原始资料写作：Claude 3 Haiku serves as the primary speculative execution model, handling the majority of queries at $0.15-0.30/1M tokens before potential escalation to more expensive models. This makes it the cost-efficiency backbone…
当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：CascadeFlow is a framework/tool that uses Haiku as part of its multi-model cascading strategy, not a standalone Haiku product deployment. The actual production impact depends on individual users' configurations and task…