There is no zero to scale to

Half the architecture of an agent backend on a serverless platform is about getting to zero. Idle timeouts. Checkpoint-on-scale-down. Queue-depth triggers. Cold-start mitigation. You build it because an idle process on a metered platform is a leak.

On a flat-rate Hoody server there is no leak to plug. Your agent workers run as daemon programs in containers, and they stay up. An idle worker burns no CPU; its container's disk blocks are shared copy-on-write with its siblings; the shared kernel caches their common binaries once. Twenty idle workers cost the server what one does, and the server costs $29 a month either way.

So you delete the scale-to-zero code. The workers are just up. The first request of the morning is as fast as the thousandth, because nothing had to wake.