I came to the exact same conclusion recently. I am building something like lightspeed that can run on “pay per second” infra like fly.io sprites because I wanted to have an open claw style agent but not have to deal with paying for another VM.
Nowadays, depending on the project’s complexity I either run just 1 sprite, several sprites, or in one case 1 permanent “CTO” VM (with a dedicated GH App identity) that specs work that a fleet of worker sprites pull and implement.
Came to a similar decision 5 months ago and ended up building “seven” on Fly’s sprites: https://github.com/1to10partners/seven
Nowadays, depending on the project’s complexity I either run just 1 sprite, several sprites, or in one case 1 permanent “CTO” VM (with a dedicated GH App identity) that specs work that a fleet of worker sprites pull and implement.