APIs in the AI era. Over time you start to notice that APIs aren’t really behaving like endpoints anymore. That mental model made sense when software was a set of deterministic pipes, you call a function, you get a result, you display it on the frontend, and everything is explicitly wired together by the developer. But once you plug an LLM or AI into the software system, the interaction pattern changes. The model doesn’t see an endpoint, it sees an interface it can reason about. It inspects the schema, forms a prediction about how to use it, tests that prediction with a call, adjusts based on the response, and folds the new information back into its internal state. It’s much closer to how a human learns a new library than how a machine traditionally consumes an API. And when you add layers like function-calling or MCP, the API starts to look like a capability surface. Something the model can navigate and integrate into a larger plan. At some point you realize APIs aren’t external resources anymore. They’re becoming part of the runtime itself.