The customer's agent waits all month for one webhook

You sell a bookkeeping agent. Each customer gets one. It wakes when a Stripe payout lands, reconciles it against their invoices, files the result, and goes quiet. For most customers that's four events a month.

On a per-second sandbox platform, an agent that exists all month to work four times is a bad deal, so you build a cold-start pipeline: spin up, load the customer's context, run, tear down.

On Hoody each customer has a container holding their agent, their ledger in SQLite, and their history on disk. An exec script receives the webhook and POSTs the event to that container's agent session. The agent runs, writes, stops. The container never left; it sat idle on the flat-rate server sharing disk blocks with 400 other customers' containers, burning no CPU.

Waiting is the agent's main job, and on a flat-rate box waiting costs nothing.