We went from "prompt engineering" to "context engineering" to "let 'er rip" in less than 2 years.
"Prompt engineering" always sounded cheesy to us "converted engineers": that's not real engineering. It's "fiddling" at best. But is it all that trivial? Probably.
Soon after came "Context engineering": The idea that working with an LLM is not about one message, a "prompt" that we send out as the singular input of our request. Rather, we should manage and "engineer" the entire bulk of the "conversation", build up a cohesive structure, compress and filter out the noise while avoiding falling into the deep pits of the "telephone game" as we take many turns to solve a complex problem.
Just a few short months later, and Ralph is taking the wheel. "Harness engineering" entered the realm. Just let the LLM do it all: let it bang its proverbial head against the wall again and again, keep track of its own outcomes and let it learn as it goes. With the added twist that the learning gains are not that of Lisa Simpsons but rather that of Ralph Wiggum. RL but it escaped the labs, and instead of steroids we put it on Xanax.
This is how we can solve all problems now. Nothing to worry about here.
