Stage 2 Runtime Architecture
One responsibility per pod. Cross-scope communication only via Kafka events or public interfaces. No shared god modules.
Swarm Orchestrator
Stateless. Horizontally scalable. Decomposes an internal issue into subtasks, dispatches them to GNN/HGT, causality, ontology, and LLM workers, then aggregates a PROV-O-traced verdict.
- Kafka task/results topics per tenant
- No pod-local state
- KEDA scales on queue lag
Agent Control Plane
Single gateway to the external enterprise LLM (Kimi K3). Enforces budget, rate limits, and circuit breakers. Adds PROV-O tracing headers. Never exposes source systems to the LLM.
- Interchangeable enterprise LLM
- Token budget & circuit breaker
- Reasoned diagnosis endpoint
Kafka Bus & KEDA Scaling
Topics partitioned by tenant and scope. Dedicated consumer groups per pod. KEDA scales each scope independently on queue lag.
- Tenant + scope partitioning
- Per-pod consumer groups
- Independent elastic scaling