cascadeflow.ai provides an open-source AI agent runtime layer that observes, scores, and enforces every step of agent execution in real time. It runs in-process with sub-5ms overhead, supporting 17+ model providers and major agent frameworks without code changes. The platform enables policy enforcement across cost, latency, quality, and compliance dimensions through eight runtime actions.
Funding
Funding not disclosed
Founders
Product
Problem
AI agents increasingly make autonomous decisions across multiple model calls, tool invocations, and handoffs, yet development teams lack visibility into these internal decision paths. Existing proxies and gateways only observe HTTP-level requests, missing the critical decisions and risks that occur inside agent execution, leaving organizations unable to enforce policies on cost, quality, or compliance in real time.
Solution
cascadeflow provides an open-source AI agent runtime layer that sits directly inside the agent process, observing, scoring, and enforcing every step of execution. The platform injects organizational policies and KPIs directly into the agent, scoring each step across six built-in dimensions—cost, latency, quality, budget, compliance, and energy—plus custom dimensions defined in Studio. Eight enforcement actions allow real-time steering: allowing steps, switching models mid-run, denying tool calls, stopping execution, retrying, requiring approval, redacting sensitive data, and serving from cache. The runtime is framework-neutral, working with LangChain, OpenAI Agents SDK, CrewAI, PydanticAI, Google ADK, Vercel AI, n8n, and others, and supports 17+ model providers including OpenAI, Anthropic, Groq, Together, on-prem vLLM, and Ollama. Integration requires only three lines of code with no rip-and-replace, and the system maintains a full per-step audit trail.
Target Audience
Primary customers are AI engineering teams and platform organizations building production agentic applications who need granular control over model costs, policy compliance, and execution quality without adding proxy infrastructure or rewriting their existing agent frameworks.
Features
- In-process runtime architecture that observes and enforces decisions where they occur, unlike HTTP-boundary proxies
- Sub-5ms overhead per step, ensuring minimal performance impact on agent execution
- Six built-in scoring dimensions (cost, latency, quality, budget, compliance, energy) plus custom dimensions definable in Studio
- Eight enforcement actions: allow, switch_model, deny_tool, stop, retry, require_approval, redact, and serve_from_cache
- Framework-neutral integration supporting LangChain, OpenAI Agents SDK, CrewAI, PydanticAI, Google ADK, Vercel AI, n8n, OpenClaw, and Hermes Agent
- Support for 17+ model providers including OpenAI, Anthropic, Groq, Together, on-prem vLLM, and Ollama
- Full per-step audit trail for governance and compliance reporting
- MIT open-source licensing with pip and npm installation packages