I've been using pi for several months with herdr. The ability to easily customise your experience via extension, with a small core, has finally made me enjoy working in harnesses. Also check out pi-blackhole (or billion context) + pi-fabric. The former basically gives you unlimited context and pi-fabric is subagents on steroids.
The new Herdr desktop has a great theming system. I haven't scratched the surface yet but it's blindingly obvious the future of software is (power) users being 100% in control of their workflow, what it does, how it does it, and how it looks.
Maybe trivial, but changing the font colors, agent status placement & colors increase my productivity 5-10%. I have more ideas like thumbnail previews, but these are specific to the project being worked on.
If you take pi and its extensions, plus herdr desktop, you have an agent that can run the way you want it to, and an app that shows the info you need in the style and way you want it to. I don't think there is one single superior workflow to dominate everything.
The downfall to this is LLMs being trained for their agents. Opus on Claude Code, GPT x.x on Codex. There are certainly strong echoes to vendor lock-in in the past if you look at all of the APIs we used to have with the social web. You could combine this with AI licensing and a national identity verification system and make the whole thing fucking miserable, with locked down hardware & software like iOS. There's a fair chance this could happen, especially in the EU. That and OpenAI + Anthropic merging are two scenarios that would suck.