Posted by karagenit 12 hours ago
All agents are ephemeral and running remotely now, and my phone has become my primary device to engage with agents.
(I'm using Ampcode and a variant of DeepSeek harness)
https://github.com/getpaseo/paseo
It works really well, has the nice code diff view which what I used to open vs code for, and it lets you drive from your phone.
All the beefy work happens on my machine at home with big hardware and .env secrets, etc. And from my phone I'm able to run multiple threads with multiple models/thinking modes.
I think a testament of how useful it is, is everybody on my team switched over to Paseo pretty much in a week.
When I read what folks write about CC and TUIs it all sounds... really, really awful and I kind of don't get why any wouldn't want to use Cursor. I'm posting this to see if anyone wants to explain how/why I might want to ditch Cursor and go all in on Claude or Codex or Pi or anything else. I'm open to being convinced!
Mostly, I just want to understand. My sense of CC is that it's great for folks who are looking to spend lots of money having 4 agents one-shot an entire application from a single prompt so that they can pick their favourite when they wake up. Is that an absurd perspective? If so, tell me how you use it.
I'm someone working on fairly significant but personal projects; I have a very clear picture in my head about what I'd like to get so I tend to write extremely dense, detail-filled prompts and go through a multi-revision planning phase before kicking off agent mode on a feature. The apps I build tend to be the result of several dozen or several hundred features developed mostly serially, and I am always the bottleneck because of that whole "knowing what I want" thing.
So, that's my usage pattern. Should I be using Claude (or Pi)?
This is probably where the disconnect is - the TUI workflow only really clicked when I started working across multiple worktrees at the same time.
It’s not all or nothing. You can still be fully in the loop, read all the code and steer each session towards the outcomes you want. You’re just doing it across multiple branches at the same time and imo it’s easier to do that through the TUI.
Then when it's all over, how do you evaluate which one to keep?
I tend to have one session that I’m actively focused on as my primary session, everything else I come back to every ~5-10 minutes to nudge them along while I’m waiting on the ‘main thread’. If my core task demands more of my time to think then I check less often.
The important thing is to find what works for you. If Cursor is giving you enough productivity then that’s awesome, carry on with it! I would recommend trying out other workflows, weighing up the pros/cons yourself and ultimately find what works for you. TUI-driven development isn’t for everyone in the same way not everyone likes to use vim and that’s okay.
Their real marketing comes/came from https://github.com/claude phenomenon. Irrespective of what underlying model you are using the claude code harness just sticks this github user as git commit author and then you have to make an effort to remove it. That is just odd.
I have a plugin for rotating between providers as quota gets used up by altering the model used by the tintiweb subagents. https://github.com/gregwebs/pi-quota-dispatcher
Note that there is another extension called pi-crew which mentiones in Readme that it was completely vibecoded, I haven't tested that one.
https://github.com/nickyreinert/yaw/
I was struggeling with multiple markdown files, not only for multiple Claude sessions, but also Copilot and Codex, accross different machines and projects with diferent purposes.
The problem: This is not only about some pre-prompt instructions, its about coding instructions for different lanugages, frameworks and architectures and skill/hooks/tools for different purposes. I drowned in multiple setups (I have ADHD/ASS, I start two projects a week). So I decided to built this kind of "wrapper" that I can use with every new project. It allows me to maintain a global setup of "defaults" and then add modules to each project as I need them.
It's not fully done now, but it already helps me. Someone else might find it useful.
My setup is Pycharm, datagrip and iTerm2.
I run all the tests, debugging, linting in the terminal and use the jetbrains gui features which I find useful (some git management basically).
It works great, and this is for building large enterprises applications.
I'm spawning a lot of agents to work in parallel. I guess I should just use the API, but it just shows that it does some random stuff you might not agree with. 99% of the users are ok with it because they don't have a 10gb repository