nught

As models getting stronger and more intelligent, complex harnesses seem not applicable anymore, less tools is better than more, less skills is better than more, less instructions is better than more. Sometimes, you don’t need to waste so much tokens on every single tasks with the existing harness, pi is the most minimal and beautiful form of it.

On other hand, making your own agents, and improve your agents seem the challenges nowadays, without the traces, observability, judges/evals, it’s hard to find a way to systematically improve your agents. We need a platform to enable experimenting on the configuration, sharing just as simple as editing the Google Docs and Scratch (slightly more technical version)