Agent API Spec Design: When API Callers Change from Application to AI Agent

In current AI Agent development, we are used to feeding context to Agents using Skills, Tools, MCP, or Prompts.

As agent tools continue to evolve, the initial hype around MCP has cooled down. While it standardizes how APIs are called, it also makes things more complex. It requires clients that support its protocol, and the server needs a matching protocol implementation. If it's just wrapping a bunch of API endpoints, why not just write code to call those APIs directly? Skill is a better alternative for this. More importantly, it provides instruction manuals that models can easily understand....

I agree that MCP adds complexity in some cases. I’ve found that using a Skill-based approach can simplify agent orchestration while keeping context manageable. Curious how others handle cross-agent communication efficiently without over-complicating the setup.