Comment by soysamurai

I got a little annoyed (okay, more than a little) with agentic developer tools and started building a replacement that I use every day now for development - both at home and at work. I have found that starting with an idea or design, using that to create a proper implementation plan, and then asking an agent to build something has been the most effective workflow for getting what I want (and that I don’t trust most models to not accidentally delete everything from my computer).

So I built an app that combines that workflow with some code dependency graph and sandboxing ideas (semi-structured build manifests with auto-installing tools / isolation) from a startup I previously co-founded, along with a handful of other tools / features (like virtualized code workspace over gRPC that works locally or remotely, starlark for scripting, a debugger with an investigation log and proposed fixes, dependency analyzer, security reviewer, and per-agent models/providers so I can use more expensive models for more complex work, like design or debugging work, and cheaper models for writing code) and built my own day-to-day development assistant that I’m pretty happy with!

As an added bonus, since I don’t like being stuck with one provider, it supports most any LLM provider (including local ones) and I can seamlessly switch between providers or models in the middle of my work if one is offline or going in circles.