Cascadeflow

cascadeflow is an open source runtime layer that runs inside an AI agent and governs each step it takes. It observes every model call, tool call and handoff, scores it against your policies and goals, and then allows it, reroutes it or blocks it, so agents stay within cost, quality and compliance limits while they run rather than being audited afterwards. It is for teams putting agents into production who need control over spend and behaviour. Each step is scored across six built in dimensions: cost, latency, quality, budget, compliance and energy. Eight enforcement actions act on the result: allow, switch model, deny tool, stop, retry, require approval, redact and serve from cache. Domain detection routes each step to a model cascade suited to the work, for example a code tuned drafter that escalates to a flagship model when tests fail, or a fast low cost model for summarising data. It is framework neutral, with support for LangChain, the OpenAI Agents SDK, CrewAI, PydanticAI, Google ADK, Vercel AI, n8n and others, and connects to OpenAI, Anthropic, Groq, Together, vLLM, Ollama and more than 17 providers. It installs from pip or npm, can start in an observe only mode with no code changes, and keeps a full per step audit trail with overhead the vendor puts below five milliseconds. The core is free under the MIT licence. cascadeflow Studio, the managed version with visual Policy and Domain builders, custom scoring dimensions and fleet analytics, has no published pricing; access is arranged through the vendor.