API governance has three layers — design-time, runtime, and consumption-time — but the industry has largely ignored the third. As AI agents become primary API consumers, inconsistent parameter naming and response shapes across APIs create real failures that humans would shrug off. The author argues for shifting governance to the consumption layer, where a mechanical engine can standardize what producers ship rather than relying on everyone agreeing to a style guide. To illustrate this, four tools are compared fronting the same two Open-Meteo APIs: KrakenD (declarative REST aggregation), Tyk (OpenAPI-native, one-upstream-per-API), agentgateway (Rust-based MCP/A2A proxy for agents), and AWS Labs' OpenAPI MCP Server (a local stdio sidecar). Each creates a different consumption contract, and the right choice depends entirely on who the consumer is. The key tradeoff is legible declarative config versus flexible but opaque code.
