The terminal should not own the work
6 points by sspaeti
6 points by sspaeti
Reading this feels like watching a child discover that the stove is hot.
I think I'm missing something here. It sounds like tmux or screen would do exactly the same with much more power.
FWIW, the list of features is considerably more detailed, and e.g. shows various integrations with Claude Code and git. I don't really have the problem that this is trying to solve, so sticking to tmux or screen makes sense to me, but...
tmix is this principle applied to shells, but this is proposing it for agent harnesses.
What makes agent harnesses unique compared to any other interactive app that's launched in a terminal?
The big feature that everyone seems to like is some kind of "waiting for input" indicator for each managed process.
But that could be added pretty easily to a tool like tmux and various agent harnesses. Maybe just use the old "beep" control code, and keep it as a sticky flag?
Between tracking of the bell, monitor-silence and monitor-activity, I think tmux has pretty good coverage already!
It's not unique. You already have plenty of daemons on your system. You don't run sshd in a tmux window (as far as I know).
No, but I do run applications like agent harnesses inside tmux for exactly the effect that this project seeks to achieve.
this reads as AI generated to me.
This matters because agentic coding is increasingly a workload rather than a conversation. Tasks wait on quotas, gates, retries, child branches, and external checks. Their lifetime should not be coupled to the lifetime of the interface that launched them. The terminal can show the work. It should not own the work.
Nearly "it's not x, it's y." along with the staccato declarative sentences. This is slop.
Nearly
Can we please not outlaw comparisons?
I'm salty I can't use emdash any more, and now any negation is sus.
AI has opened the door to realizing so many unoriginal ideas which solve solved problems, teaching the prompters little else but shortcuts.
Its seems do a lot more than my first impression, after reading more into it. Automates headless cli agents, to do all sorts of agentic work patterns and workspace setup and management in a repeatable way.
It looks like it has an API. I may try it. I learned that there are some other projects trying to solve the same problem.
Isolated worktrees, sandboxing, specific harness settings per project, etc, is a pain point. And other solutions push into the cloud or into an Electron App. So, vincent's concept seems interesting as it could provide a lot more control, transparency and progammability versus other paths.