The very nature of LLMs means you can't invent a thing for current agents to use that they'll be better at using than the things they already know how to use from their immense training data. You can give them skills, sure, and that's useful, but it's still not their native tongue.
To make a thing that's really for agents, you need to have made a popular thing for humans ten years ago, so there's a shitload of code and documentation for them to train on.
we have a custom .yaml spec for data pipelines in our product and the agent follows it as well as anything in the training data.
while I agree you don't need to build a new thing "for agents", you can get them to understand new things, that are not in the training data, very easily.
What makes something like this "for agents", anyway? It's opinionated...a human's opinions, I assume, since agents don't want anything and thus can't have opinions. But, many existing tools are opinionated. Types are good for agents, because it keeps them honest, but many existing things in this space have types. Python is good for agents, because there's a shitload of Python code and documentation in their training data, but many existing things are built with Python (and TypeScript, Go, and Rust are also typed languages and well-represented in the training data).
I dunno. I think a lot of folks are sitting around with an agent thinking, what can I build? And, a lot of things "for agents" are being built, as a result. I think most of them don't need to be built and don't improve software development with agents. They often just chew up context and cache with extra arbitrary rules the agent needs to follow without delivering improvements.
Everything which just works "by convention" or by "opinionated defaults" (allowing a tightly coupled but very feature rich framework) helps to reduce the noise / lines that needs to be reviewed.
While this approach might not be optimal for every project, I'm certain the opinionated defaults can work for many endeavours. And the reduction of complexity might be one important aspect, which can make an "agentically engineered" project sustainable.
This is exactly why I've gone back to Ruby with Sinatra or Rails for my personal side projects, despite Ruby's horrid performance.
As long as you are content to remain on e.g. Rails' "Happy Path", then I've found agents do a fantastic job because there's lots of Ruby in the training set and there's less surface area where a context mismatch/hallucination can end up going off the rails. Pun only partially intended.
Django code is pretty easy to review quickly. LLMs are good at writing it.
Django is just old and bloated, so the fork is a good idea. Maybe I will use this for my next side project.
- fork of django
- it's opinionated
- typed
- comes with skills / rules / docs baked in
I'm not against this idea in principle, but I'm also not sure why that is better than what's already out there, except maybe you save some tokens by not vibe coding this yourself?
I do think in the future we'll see some novel libraries that are agent-optimized first. I'm not sure if this is it, though.
(edit: formatting)