AI Agent Orchestration Platform · View 05 of 32 · 2 · Structure
Decisions
- Control plane on AKS: long-lived, moderate scale, needs fine-grained network policy and workload identity
- Execution plane on Container Apps: KEDA scaling on queue depth is the whole requirement, and a managed environment removes a cluster to operate
- Code and browser tools run in Container Apps dynamic sessions — Hyper-V isolation per session, destroyed after use
Sizing basis
- 20 to 400 runtime worker replicas covers 5,000 concurrent executions at an 8× peak-to-average ratio
- Service Bus Premium at 4 messaging units sustains the dispatch rate with headroom for retry storms
- Cosmos DB provisioned with autoscale; the checkpoint write is the dominant RU consumer, not reads
Risks
- Cold-start latency on scale-out from low replica counts; mitigated by a minimum warm pool on the interactive queue
- Dynamic session pool exhaustion under a code-heavy workload; needs a per-tenant session quota before general availability
- AKS and Container Apps means two runtime operating models; accepted, revisit if Container Apps gains the network controls the control plane needs
