@petergyang In general, these harnesses put too much onus onto the user. It’s nice to see the wiring as a power user (for now) but it’s very much against standard design principles.
@petergyang I prefer locally using launchd because then I delegate the crons to various harnesses versus being locked into one. Cleaner and more control
@nikunj Humans are inherently lazy. While it obviously wastes more time, it’s easier on the brain to have many of the same and feel busy. It’s madness.
@thsottiaux 3p model support. If you easily supported 3p models like Claude and Cursor, I would use more Codex. Instead I end up using a lot of tmux in the bg
@keeran@pvncher@yacineMTB I run them on haiku/mini for simple tasks. My orchestrator agent decides what sub agents to spawn and it actually becomes more token efficient
Do Trump Account tax benefits suck? Example: put $10k in a standard brokerage account, it grows to $100k by the time your kid turns 18, and they withdraw $40k. That could be taxed at 0% long-term capital gains. In a Trump Account, that same $40k may be taxed as ordinary income.
@kunchenguid@myfirstmate FINALLY, someone who gets it. I’m the same, I try to use vanilla bash too. And I want to be very model agnostic.
I do find that you need to strike a good balance between working with the harness vs homegrown solutions.
@kunchenguid To clarify, my subagents are hybrid tmux and harness-specific. It depends on the use case for persistence. For adversarial review, tmux, since separate harness. For many cases, direct subagents. I think hybrid wins and the agent decides when to use it.