A lot of current agent systems feel like monoliths: huge prompt surfaces, hidden state, vague permissions, and too much “magic”. That’s fine for demos, but once an AI can touch your terminal, files, browser, and accounts, I think that model starts to break down.
The idea I’m exploring with Jeriko is simpler:
keep execution on the user’s own machine
expose small machine-level primitives
compose workflows instead of hiding everything inside one giant agent
make permissions and boundaries more explicit
So the thing I’m really testing is not “can I make an AI OS”.
It’s more:
does Unix philosophy map to AI at all?
Meaning:
should AI systems be built more like small composable tools?
are CLI/process/file/browser primitives the right abstraction?
does nondeterminism make Unix-style boundaries more important, or does it make the analogy fail?
I’d especially love feedback on 3 things:
whether this framing is actually useful or just romantic Unix nostalgia
what permission/sandbox model you’d expect before trusting something like this
whether local-first AI tooling is meaningfully better than hosted agent products, or just harder to build
Happy to answer technical questions and criticism.