Cloud AI platforms¶
Each cloud's AI platform is built to call its own models, and much of it can't be pointed anywhere else. What can go through the gateway is worth knowing exactly: there, the calls are checked against your policies, charged to the right team and recorded in Request logs beside everything else your organisation sends.
| Platform | Through the gateway | Not through the gateway |
|---|---|---|
| Azure AI Foundry | Prompt agents using a model connection to the gateway; your own code | Agents using models deployed in Foundry; features Microsoft lists as not working with connected models |
| AWS Bedrock Agents | Your own code: action-group Lambdas, agents you host on AgentCore | A Bedrock agent's own orchestration (always a Bedrock model) |
| Vertex AI Agent Builder | Agents you write (ADK and others) and deploy to Agent Engine; your tool code | No-code agents built in the console |
| Databricks | An external model endpoint pointed at the gateway: ai_query(), the AI Playground, your agents |
Genie and Databricks Assistant |
The rule behind every row: when your code (or a setting you control) chooses where a model call goes, it can go through the gateway; when the platform calls its own models for you, it can't.
The gateway reaches these clouds' models as well (Azure OpenAI, Bedrock, Vertex AI), once an admin adds them as providers. Moving a call into the gateway doesn't mean leaving a cloud's models, only bringing them under your policies.
In every case the platform calls the gateway from the cloud's own network, so the gateway must be reachable from there.