I cancelled my subscription when it got bought by space nazis. But I actually use the cloud agent flow cos it's amazing. Subsequently I built my own software factory based loosely on the OpenAI Symphony pattern.

I used to be a heavy shell and IDE user and now I ask my local AI to do the majority of computing tasks I used to do by hand.

Codex/similar harnesses allow me a digital workspace where I can organize both short-term and long-horizon tasks and build a "journal" of my computational history. I journal less and talk with my agents more. When I need to remember something, I can Ctrl+K in my harness or ask my agent to search previous conversations.

In Codex in particular, each workspace is unassuming, offering several panes with various functionality that I can customize to suit each particular task, and Codex remembers the layout for each conversation. You can structure some tasks to offer an IDE-like workspace, while others can just be a simple conversation pane. Cursor feels ham-fisted in comparison with non-coding tasks.

If I need to rename a few files, that's a job for my agent. If I need to convert something with ffmpeg, I'm never looking up reference again, because that's a job for my agent. If I need to change more than a word or two in a line of code, that's a job for my agent.

My job is to build a conceptual map of my work environment so that I can understand how each interface communicates without needing to know the exact implementation (and I still do quality/security review passes). I plan out work with the agent, design acceptance criteria and then monitor my agents as they work on various tasks across several projects. The agents' ability to design bespoke diagnostic programs, use a browser, etc. have allowed me to offload some amount of QA testing as well. I still have to test and review, but I don't have to write regressions by hand anymore or test for the more obvious UI/UX issues.

It's not a perfect system but from top-to-bottom, my computing experience has radically changed in a relatively short period of time. Putting this level of trust into cooperative agents was simply not possible a year ago, and not even six months ago. Harnesses have gotten better, and the last few rounds of cutting edge models in particular have really changed the landscape.

I'm sure I will grow to miss writing code by hand, some engineers have already expressed this to me, but I also have RSI to manage and honestly these agents have been a godsend. My wrists hurt less than ever after a long day of working, and yet I'm working on more things at once with a higher velocity without sacrificing architectural quality, security, robustness or the other big reasons why I previously could not trust agents.

Of course, like many others have expressed, this new paradigm is a lot more exhausting. That's also a good thing, because I have lifelong chronic insomnia. I'm still able to care for my mind and body more as I'm able to take more frequent short breaks to stretch, walk around or clear the cobwebs out of my brain as I know that progress is still being made. And I can still work from my phone while I walk.

Devin Desktop fka Windsurf

> how annoying it was to try to cancel Cursor

Add it to the enshittified software list.