Stop prompting AI agents one instruction at a time. Design loops that drive them, and learn how to stack loops upward as models improve. This is a translated excerpt from Latent Space / AINews on agent orchestration.
This is a translated and adapted excerpt from Latent Space / AINews.
There is a lot of “loop discourse” in the air:
- Steipete: “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”
- Boris: “I don’t prompt Claude anymore. I write loops, the loops do the work.”
- Andrej on Autoresearch: To get the most out of the tools that have become available now, you have to remove yourself as the bottleneck. You cannot be there to prompt the next thing. You need to arrange things so they are completely autonomous, maximize token throughput, and not be in the loop. The goal now is to increase your leverage. You do not want to be the researcher in the loop looking at every result and holding the system back. The question becomes: how do I refactor the abstractions so I arrange it once and hit go?
This framing is compelling, and many people do not realize how many loops we are already inside:
A more minimal set of loops:
One could argue that the entire game of the next century is learning how to stack loops as effectively as possible. In the early days of each phase, it is valuable to know when to go down a loop when things go wrong, for reliability. But as models improve, it may be even more valuable to know how to go up a loop, for leverage.
If you do not figure out how to do this, do not be salty when you lose to those who do.
Rich has his “Bitter Lesson” for models. We now have the Salty Lesson for agents:
Do not fix things yourself, as you have historically done.
Instead, focus on systems that scale with more agents, such as goals and orchestration.
