> ## Documentation Index > Fetch the complete documentation index at: https://docs.cascadeflow.ai/llms.txt > Use this file to discover all available pages before exploring further. # Why cascadeflow > The business and technical case for inside-the-loop agent runtime intelligence. # Why cascadeflow cascadeflow is a **library** and **agent harness** — a Python and TypeScript package you install and import, not a proxy or hosted service. It runs in-process inside your agent code. Most AI optimization tools sit outside the agent — at the HTTP boundary, in a proxy, or in a dashboard after the fact. cascadeflow sits **inside** the agent loop, where decisions actually happen. This is the difference between watching outcomes and controlling them. ## 1. Inside-the-Loop Control Is the Core Moat cascadeflow influences decisions at every agent step: model call, tool call, sub-agent handoff. This is where most cost, delay, and failure actually happen. External proxies only see request boundaries. cascadeflow sees **decision boundaries**. ```python theme={null} # Every step inside this run is governed — not just the HTTP request with cascadeflow.run(budget=0.50, compliance="gdpr") as session: result = await agent.run("Process EU customer data") # Budget tracked across 12 tool calls, 3 model switches, 2 sub-agent handoffs ``` ## 2. Multi-Dimensional Optimization Creates Enterprise-Grade Value Most tools optimize one metric — usually cost. cascadeflow optimizes across **six dimensions simultaneously**: | Dimension | What it controls | Who cares | | -------------- | ----------------------------------- | -------------------- | | **Cost** | USD per call, per run, per user | Engineering, Finance | | **Latency** | Wall-clock ms per call and total | Engineering, Product | | **Quality** | Model quality priors and targets | Product, QA | | **Compliance** | Model allowlists (GDPR, HIPAA, PCI) | Legal, Security | | **Energy** | Compute-intensity coefficients | Sustainability, Ops | | **Budget** | Cumulative spend caps and limits | Finance, Engineering | This makes cascadeflow relevant not just to engineering teams, but to finance, security, operations, and sustainability stakeholders. ## 3. Business Logic Injection Turns AI from "Smart" to "Governable" Organizations can embed KPI and policy intent directly into agent behavior at runtime. This shifts AI control from static prompt design to **live business governance**. ```python theme={null} # Business intent encoded directly into agent behavior @cascadeflow.agent( budget=1.00, compliance="gdpr", kpi_weights={"quality": 0.7, "cost": 0.2, "latency": 0.1}, kpi_targets={"quality": 0.9}, ) async def eu_premium_agent(query: str): return await llm.complete(query) ``` ## 4. Actionability Is Immediate, Not Advisory cascadeflow does not just observe and report. It can **directly steer runtime outcomes** based on current context and policy state: | Action | What it does | When it triggers | | -------------- | -------------------------- | --------------------------------------- | | `allow` | Proceed with current model | Policy checks pass | | `switch_model` | Route to a different model | Quality, compliance, or KPI mismatch | | `deny_tool` | Block a specific tool call | Tool cap reached or risk policy | | `stop` | Halt execution entirely | Budget exhausted or hard constraint hit | This closes the gap between analytics and execution. ## 5. Transparency De-Risks Enterprise Adoption Every runtime decision is traceable and attributable. This supports auditability, faster tuning cycles, and trust in regulated or high-stakes workflows. ```python theme={null} for record in session.trace(): print(f"Step {record['step']}: {record['action']} — {record['reason']}") # Step 1: allow — budget ok, compliance passed, quality 0.92 # Step 5: switch_model — quality below target, switching to gpt-4o # Step 8: stop — budget exceeded ($0.50/$0.50) ``` ## 6. Latency Compounding Is a Structural Advantage Proxy-based optimization adds 40-60ms per model or tool call from extra network hops. In agentic workflows with 10+ calls, that creates **400-600ms of avoidable overhead** per task — and much more for deeper loops. cascadeflow's in-process approach adds \<1ms per call. Optimization does not come with a latency penalty. | Scenario | Proxy overhead | cascadeflow overhead | | --------------------------- | --------------- | -------------------- | | Single call | 40-60ms | \<1ms | | 10-step agent | 400-600ms | \<10ms | | 25-step deep loop | 1-1.5s | \<25ms | | Real-time UX (100ms budget) | Consumes 40-60% | Consumes \<1% | This is critical for real-time UX, task throughput, and enterprise SLA performance. ## 7. Value Proposition Is Measurable and Defensible cascadeflow proves impact with reproducible metrics on real agent workflows: ```python theme={null} summary = session.summary() # { # "cost_total": 0.0847, ← actual spend # "steps": 12, ← agent steps taken # "tool_calls": 8, ← tool executions # "budget_remaining": 0.4153, ← governance headroom # "energy_used": 34.2, ← compute intensity # } ``` Better economics and latency while preserving quality thresholds — not a trade-off, a structural improvement. ## 8. Why This Can Become a Category Leader * **Framework-neutral and provider-neutral** — works with LangChain, OpenAI Agents, CrewAI, Google ADK, Vercel AI, n8n, Hermes Agent, and custom frameworks * **Solves a structural gap** orchestration frameworks are not built or incentivized to solve * **Expands from optimization into business-intelligence control** for agents * **In-process architecture** is fundamentally better than proxy architecture for agent workloads ## Strategic Outcome cascadeflow can become the **default intelligence and governance substrate for agents**: the layer companies rely on to make agentic systems economically viable, policy-compliant, and operationally predictable at scale. ## Start Now pip install cascadeflow — start observing immediately. Every feature, with links to GitHub examples.