Connect a Postman API key to check your work
You do each step in Postman, then LiftOff uses your key to look at your workspace and confirm it's done. Connect once and it works for every module. Your key is saved in this browser and is only used to check your work in Postman.
Review the governed support-agent architecture you assembled, the differences you observed between Direct and Fabric, and the practices to carry into a production implementation.
Review What You Built
You started with a working fintech support agent that held separate connections to its model, Banking MCP server, Support MCP server, and Fraud API. Without changing the agent's orchestration logic, you placed Fabric Gateway between the agent and those upstream services.
Your completed Fabric path
| Capability | What you configured | Result |
|---|---|---|
| Model access | An LLM service and route with its provider credential stored in Fabric's managed vault | The agent calls the model through a governed Fabric URL without holding the provider key |
| Fraud assessment | The Fraud API imported from its OpenAPI definition, plus a managed service credential | Fabric supplies the upstream API key and exposes the API through the data plane |
| Banking and Support tools | Two authenticated MCP services exposed through one route | The agent manages one MCP connection instead of two |
| Tool discovery | Tool Search on the combined MCP route | Fabric replaces 103 complete MCP definitions with search_tool, get_tool, and execute_tool |
| Model governance | An LLM Rate Limit instance on the model route | Fabric enforces a token budget and returns a 429 when the lane exceeds it |
| Adaptive model routing | An Adaptive Model Router instance backed by a decision service | Fabric can match model capability and reasoning effort to each prompt's difficulty |
| Operational evidence | Analytics and Traces | You can examine aggregate usage and investigate individual requests and policy decisions |
What the comparison demonstrated
- Direct begins with 104 tools and sends their definitions to the model without passing through your Fabric policies.
- Fabric Gateway begins with four tools—the three Tool Search discovery tools and the Fraud function—while retaining access to the underlying Banking and Support capabilities.
- The smaller Fabric catalogue uses fewer tool-schema tokens, less context, fewer model tokens, and typically costs less to run. Exact measurements vary between runs.
- A policy added at the gateway changed the Fabric lane's behavior without requiring a change to the support agent or model provider.
- The Direct lane remained unaffected, making the side-by-side run clear evidence that Fabric enforced the 429 response.
Key takeaways
- Fabric Gateway does not replace the agent or its orchestration. It gives that orchestration one governed path to the services it needs.
- Managed credentials keep upstream secrets out of agent code and client requests.
- A single route can aggregate multiple MCP services, while prefixes preserve each tool's source and prevent naming collisions.
- Tool Search progressively discloses capabilities instead of loading an entire tool catalogue into every model turn.
- Adaptive model routing can reserve more capable models for harder work while sending straightforward requests to less expensive configurations.
- Plugins apply policy centrally, and Analytics and Traces provide evidence of what happened at runtime.
Clean up your workshop configuration
- Disable or remove the temporary
workshop-llm-token-limitinstance if you have not already done so. - Treat the public MCP routes used in this workshop as a learning configuration. Use gateway authentication and access policies appropriate to your users and environment before deploying a production workload.
- Keep provider, API, MCP, and gateway credentials out of source control. Rotate any credential that may have been exposed.
Expected outcome: You can explain the end-to-end architecture, identify where Fabric applies credentials and policy, and describe why the governed Fabric lane uses fewer model resources than Direct.