I've shipped all four. I just didn't have the buzzwords for them.
I built real-time voice agents for dispatch — that's a harness. A policy-search RAG stack — that's context engineering. A multi-modal drone-detection system running its own observe-decide-act cycle — that's loop engineering. Nobody called them that. We called it the system.
The trap with the names
Prompt engineering had courses and job titles before context engineering buried it. Harness landed before most people finished learning context. Loop landed before the ink dried on harness. The skill you grind for months now expires in weeks — obsoleted not by your competitor, but by the next model that no longer needs the trick.
The question isn't "have you learned the latest layer." It's: which layer survives the next model?
Sort by durability
Prompt is already half-commoditized. Every release eats another bag of tricks.
Context is the fragile one. RAG, chunking, vector DBs — most of it is scaffolding we built around one fact: attention is quadratic, so a model can't hold a whole codebase or contract in its head. A startup called Subquadratic is betting that fact is optional. If linear attention is real, a whole layer is dead weight. Big if — the same demo is being called the biggest breakthrough since the Transformer, or AI Theranos.
Harness and loop are the durable ones. A smarter model still calls tools with bad arguments — it still needs bounded, reversible execution. And a model that's outgrown your prompts needs you out of the loop, not in it. That's Karpathy's whole point: stop being the bottleneck.
The salty lesson
The bitter lesson was for models. The salty lesson is for us — stop fixing things by hand, build systems that scale with more agents.
Don't bet your next two years on a layer the next model deletes.
Or are we all just speed-learning scaffolding for a bottleneck that's about to disappear‽