I became interested in how the harness works internally so I could try to implement my own small AI agent in a business project. And I realized that right now, most of the model’s efficiency lies precisely in the harness. And that there’s a whole complicated world inside it, one that will soon become more complex than the models themselves.

All my experiments using third-party harnesses for code-writing agents showed that if a model is used in a non-native harness, it sometimes burns through 5 times as many tokens, and the results are worse. That’s why I now use MCP calls not to the models directly, but to new sessions within their harnesses.

Now, almost every day I see new announcements about coding agents with bold claims about their superiority. But in none of them have I seen a clear description of how they’re better or exactly how their architecture for working with models is superior to the original harnesses.

As I understand it, Pi is like "Lego" for building your own harness. But do you know what happened after I’d been using real Lego for half a year? I have 15 kg of it loose in three bags. My son and I have almost given up on the idea of sorting it out and actually using it properly. It’s pretty much the same story with my MCP, plugins, and skills for Claude.