Building with agents
I’ve been building pieces of AI agents almost every day lately. There’s been a lot of trial and error: different architectures, different models, different instructions, different context. Mostly trying things, seeing where they break, and slowly getting a better sense of how these systems should be put together.
A task might start with one model, move to another, branch into a few different approaches, come back for comparison, an eval, a judgment, another iteration. It feels much less linear than the way I’m used to working. At any given point there are several paths worth exploring, and now I can actually explore most of them. That has changed what feels scarce.
I can read more, test more assumptions, follow more rabbit holes, and try more ideas in a day than I could before. Getting another answer is easy. Getting twenty reasonable answers is also pretty easy. Figuring out which one is actually valuable is not.
So I find myself spending more time deciding what matters, what’s missing, which direction deserves another round, where another thousand tokens might actually help, and when it’s time to stop. There’s a strange new discipline in having almost unlimited opportunities to keep going.
And of course, humans hallucinate too. We anchor on the first good idea, overlook things, get attached to elegant explanations, and sometimes confuse confidence with being right. Having models around doesn’t remove any of that. In some ways, it gives you more opportunities to do it.
The more I work this way, the more I suspect that judgment and cognitive bandwidth are becoming the real bottlenecks.
LLMs let me think across a much larger surface area than I could on my own. The hard part is still deciding where to look.