cascadeflow

Why we included this project

Most of an agent's spend goes to repeated model calls inside long loops, and the usual fix is a proxy parked in front of the API. cascadeflow works in-process instead: small, cheap models answer the straightforward steps first, a quality check decides whether the answer holds up, and only then does the call escalate to an expensive flagship model. Around that routing, a harness enforces budgets, tool-call caps, latency and energy limits, and compliance rules, with KPI-weighted routing and per-step decision traces for auditing. It plugs into LangChain, OpenAI Agents, CrewAI, PydanticAI, or n8n without swapping providers, ships in both Python and TypeScript, and the README reports savings of 69% on MT-Bench and 93% on GSM8K while keeping 96% of GPT-5 quality. Teams whose traffic is mostly easy queries a small model can handle will get the most out of it, since that's exactly the pattern the cascade is built to exploit.