@petekp@linear Interesting, this is one micro-interaction I never noticed myself (for some reason I’ve never felt like expanding all toggles)
I wonder how consistently it’s implemented
What if we treated software development like the R&D activity it actually is, rather than the manufacturing activity that we’ve mistakenly thought it was for decades?
software factories feel like a temporary problem. the models will learn to build them for you
the more interesting question to me is how to build a workspace where agents can collaborate with you
@martin_buur I found out recently that the free plan of Muse has got a lot more usage than even my paid Grok Bot (or so it seems).
Reminders and stuff is what I use the free plan for there.
Now I’m reserving Grok Bot for “smarter” things I need done like design, dev, sysadmin chores.
I think it’s a good thing that all of the world’s interaction design & UI engineering talent gets diverted from building CRUD apps, e-commerce apps, and FinTech apps that are better handled by natural language commands; and more towards “direct manipulation” creative apps.
In this AI future, I’m not sure we’ll be building UIs as often as in the past. Most folks will prefer to interact with their AI (via voice or text) instead of browsing websites to get things done.
We’ll still build services, but just won’t need to build the visual app UI.
I also think a cheaper model (than Opus 5.5) can’t do this well.
I tried doing this with Grok Bot and Muse, both of which support artifacts.
The same skill produces terrible artifacts! So, for better or worse, I don’t mind this way of working.
I also already work with annotations in my UI (I use agentation), so this is more for higher-level calls involving UX, architecture, DevOps, etc.
I have Claude write to a “decision kit” artifact per thread as a summary of what it needs from me (UI or architecture to review, decisions, taste calls).
It’s not deterministically enforced by the harness so sometimes it does it poorly, forgets to update it, creates duplicates, etc
But most of the time it works, and artifacts show up as first-class elements in the Claude desktop app UI linked to the threads that produce them
So now I work at the level of artifacts instead of threads. I used to pin threads, now I just pin artifacts.
It is token hungry, though, and I use up an Ultra plan in 4 days.
I do this with artifacts in Claude Projects, and it works pretty well (albeit with minor rough edges in the UI).
Each thread has one or more artifacts for the state of the work: decisions, things to review, that is updated as the work evolves.
I just work at the level of artifacts now. I used to pin my WIP threads. Now I pin artifacts instead.
The UI’s also been updated now to show artifacts next to threads which is super nice, and every artifact has a backlink back the threads that they were created/edited in!
So if I need to change something it’s a prompt away.
Still a little rough so I’d give it until it’s out of beta!
But it’s super nice that you have a coordinator thread that can manage all other threads, and have multiple threads that can talk to each other, and have working memory that’s outside of the repo.
But it’s largely cloud-based and if you like to work locally it’s a bit of a pain.
It doesn’t help that:
1. Persistent agents are designed to run on their own unattended, which means it’s easier for even non power users to max out their usage limits
2. Third-party models (outside of Grok models) have less of a cost advantage, and they started utilizing them recently
3. Unification of X, Grok, Cursor reduces revenue per user, for users who are sub’ed to >1 service
This could easily become sustainable when they have a better Grok model that rivals Opus 5.5. Maybe 4.8?