Experimenting with Agent-Native Architectures

Note: This is just stream of consciousness (dictated) as this is a post on a short-term topic that'll likely just get taken down at some point.

I've been building with Claude Code, like many have, and thought that the recent every posts and writing on agent-native architectures were interesting, mainly trying to paradigm shift to building quickly, effectively Claude Code in a trench coat style. Just learning about this and experimenting with it, I don't mean to endorse it.

I found in past exercises that when trying to do something highly agent-native with software, I typically run into the classic software problems that I'm solving, which means better separation of concerns, cleaner scoping, et cetera. It usually leads me more away from LLMs and agents than towards it. However I think this is a super powerful way of building especially if you don't need scale or have something ultra complex. I also think that the power and cost of inference obviously is just going down a ton, so it makes a ton of sense to begin using these primitives even if a bit expensive currently.

What I've done is built a relatively small scale search service, call it up to 50 documents in a folder or directory. I think there are a ton of search problems where this scale very much suffices. That works super well, and the agent-native approach was really performant for me. Additionally, in the work I'm doing, I've got some other one-off services where I flipped into an agent-native approach and found it pretty performant.

I haven't played a ton with robustness cost or optimization. We'll see if this withstands the test of time, but it's a cool new pattern and paradigm shift. Not sure if there's something I do believe that the way that we build and our building blocks and primitives are changing for sure. I'm not sure if this is necessarily the most durable pattern or not. Even some stuff that Evry has done seems pretty experimental, but cool to try out and appreciate the team building openly such that others can use their components and approaches.