My interruptions and frustration with llms has gone down a cliff. It’s amazing to see what just a few lines of system prompt really can do to keep agents doing the correct thing much more often.
So there are many wonderful dev experiences, but generally relying on something relatively simple like pi makes a lot of sense, for my use cases.
Obviously you can add skills/rules/prompts/etc, but I don't find those are particularly great at being stuck to.
Pi is much more minimal, so if you have strong opinions about how you want to work with LLMs, you'll have a much better chance of bending the to your will.
But with Pi being barebones and intended to be customized, there are a million extensions to choose from compared to OpenCode. This means not only can you write your own extension easily, 5 other people already have. I forked one popular question-asking tool and added support for an external ui (Zenity). And an LLM provider (Synthetic) had a bug which I helped get fixed in a 3rd party extension for the provider. Both I got done in days, because I didn't have to go through some official Pi team; extensions distribute the work better. It's still hard to manage all the extensions, but better than the alternative.
You can do this with any harness. I really dislike the Node.js ecosystem and avoid Pi and the like. I started with another one written in Go that has all the basics and I can extend it in minutes for anything I need while keeping track of upstream.
Another one written in Go is Crush
I wrote Juggler (https://github.com/juggler-ai/juggler) because I liked Pi's approach to everything being customisable and controllable, but I wanted a slick, navigable GUI rather than terminal hell.
...BUT: having gone to great pains to build an architecture around customisation, where every possible pluggable thing is a plugin, I'm starting to feel jaded on the whole "we need to customise it all" mentality. The models have got so good over the last year that I just don't think that tweaking the harness is something many people will need to do as things progress. The models only need a bare-bones set of tools to do everything they need, and I'm starting to think that UX and UI are the only really important things to focus on in an agent.
(Although from the perspective of coding a harness, making things plugins and having a plugin API does help to keep things clean and stops the codebase getting sloppy)
There is contest.
With minimal work, you can get DSH web to be exposed to your internet behind Cloudflare Access (or similar) and have a harness running on your computer at home from any browser, including your phone's.