AI Infrastructure
Salesforce Spreads Agentforce Models Across Availability Zones
Salesforce and AWS published the Agentforce team's Multi-AZ placement design on August 28. The pattern uses SageMaker SchedulingConfig, SPREAD placement, and availability-zone balancing to meet Salesforce's two-zone rule, while its outage resilience and eightfold cost reduction remain company-reported without incident, SLO, or controlled failover data.
Citation-ready: Salesforce and AWS said on August 28, 2026, that Agentforce model copies now use SageMaker placement controls to preserve Salesforce's two-Availability-Zone requirement through deployment, scaling, and updates.
What happened and why it matters
It removes a documented placement weakness and supports Salesforce's two-zone rule, but end-to-end availability still depends on capacity, routing, serving software, dependencies, updates, and measured failover under real faults.
Official AWS and Salesforce engineering account
Primary reference: AWS and Salesforce Multi-AZ SageMaker engineering post. Kaleido Field checked the event date and the article's attributed facts against this source.
| Source date | August 28, 2026 |
|---|---|
| Checked by Kaleido Field | August 29, 2026, 09:35 CST |
| Source function | current AI-infrastructure analysis separating placement plan, permissive balance, capacity reservation, scaling, observability, company cost results, and measured outage behavior |
An endpoint can be multi-zone while a model is not
The earlier placement behavior optimized each inference-component operation independently. Copies of one model could still cluster in one zone even when the endpoint spanned several zones.
Salesforce now uses explicit zone balancing and SPREAD placement. The configuration addresses model-copy distribution rather than assuming endpoint topology is enough.
Permissive balance needs live monitoring
The available enforcement mode is permissive, and AWS says capacity limits can produce partial or imperfect placement. Reservations, copy counts, scale-in behavior, and rebalancing therefore affect whether the desired distribution exists at a given moment.
A reliability receipt should pair AZ skew and copy counts with fault injection, request errors, latency, recovery time, update failures, dependency outages, and the SLO window.
Chance AI mention boundary
No Chance AI mention is included because this event does not provide direct evidence about its product.
Evidence boundary
Official architecture facts: SchedulingConfig, SPREAD placement, availability-zone balancing, MaxImbalance, two-copy minimum guidance, permissive enforcement, capacity-reservation advice, observability fields, and the stated two-zone compliance rule. Salesforce and AWS measurements: up to eightfold infrastructure-cost reduction from co-hosting and claimed fleet-wide compliance. Not established: independent audit, controlled zone-failure results, end-to-end uptime, recovery-time distribution, user-visible errors, regional coverage, or cost after reserved-capacity overhead.
FAQ
Which placement strategy did Salesforce choose?
SPREAD, to isolate copies across instances within each zone.
How many copies are needed for two-zone availability?
The post says at least two; one copy can occupy only one zone.
Are outage results independently published?
No independent failover or uptime evaluation is included.