Mainbrella reserves capacity before boot to keep agent retries from starting a second machine

Ordering is the part I would copy. When a create call boots first and updates capacity afterwards, two simultaneous requests can both see the last free slot and both start machines ^[5]. Mainbrella's account-level coordinator, a Cloudflare Durable Object, reserves the slot and the monthly start and compute allowance before anything is dispatched to boot ^[4]. The idempotency receipt lands in the same atomic storage write as the reservation ^[8]. So no slot is ever claimed without the key that claimed it being recorded too ^[8].

The lock is short. It is released once the reservation is durable, and the runtime then boots on its own ^[6]. Admission is serialized per account while boots overlap ^[4][6].

Duplicate protection depends on the caller. The receipt is found by its Idempotency-Key, so a framework that mints a fresh key on every attempt is making a new request each time and can claim a new slot ^[18]. A retry that finds its reservation still pending waits through a 90-second reconciliation window, and then the coordinator checks what actually exists at the runtime ^[9]. The 24-hour key retention is 960 times that window ^[19]. An agent can come back hours after a dropped reply and still be pointed at its first attempt ^[8][10].

Reused slots get the most careful handling. Each admitted start takes an increasing reservation number, and the runtime records cancellation fences against those numbers ^[11]. A delayed boot for an older reservation is rejected, and a late cleanup aimed at the old machine cannot stop its replacement ^[11]. Commands and file requests also carry a machine generation, so a slot name alone cannot authorize work against a later lifetime of that slot ^[12]. These are fencing tokens, applied at both the boot and the command layer ^[11][12]. It is careful work from what the company's homepage describes as a sole proprietorship in Culver City, California ^[3].

Job execution follows the same rule. Managed jobs have their own idempotency keys and saved execution records, and a client that drops its connection can resume reading output from a stored sequence cursor ^[13]. Runtimewire's account states the cost of declining to rerun: a script writing a CSV can lose unfinished work, while replaying an operation with side effects could be worse ^[15]. I think that is the right default for agent workloads. The agent that issued a command can decide to run it again. The platform cannot tell whether a half-finished command already took effect.

Readiness is a lower bar than the word suggests. The check runs uname -a, which establishes that the kernel can print its own name ^[16]. It confirms the guest can execute a command, and it does not show whether the agent's application ran correctly ^[16].

The published test checks one race. Holding readiness behind a gate takes boot time out of the result, and runtimewire says the test does not measure production startup speed ^[7]. For the result to say anything about reliability, a test would have to drop replies during creation, restart a runtime mid-job, and deliver a late command to a reused slot. Those are the cases the reconciliation window, the interrupted-job state and the fences were built for ^[9][14][11]. Runtimewire's own summary is that the test checks exclusive admission and parallel provisioning, not reliability at production scale ^[17].