An agent workload does not look like a classic web API: a run can take a while, wait for an approval, call several tools and start sub-agents in parallel. Navin Enterprise AI therefore keeps its runtimes separate so that each one scales at its own pace: agents, real-time voice, batch and media workers, periodic tasks and inference. No runtime holds governance data.
Autoscaling on load and on events
Workloads scale with the Kubernetes Horizontal Pod Autoscaler and with KEDA, which triggers scaling from events. Event triggers (brick C10) run on NATS JetStream, and the message bus between channels and the engine works in memory or on NATS. Multimodal batch and media workers also pick up their jobs from NATS.
Durable execution: no lost work
Every step of a run writes a checkpoint to PostgreSQL and a heartbeat. If a process goes down, the AgentLoop resumes the run from its last checkpoint, and the server worker (python -m eai.worker) spots stale heartbeats, ticks the scheduler, flushes the event outbox and expires quota reservations. The SubagentManager runs sub-agents in parallel and feeds their results back into the parent turn.
Quotas and GPU accounting
Scaling out must not mean spending without limits. Before every model call, the cost is estimated and a quota reservation is taken; after the call, token accounting turns the Usage Event into a measured cost and settles the quota. Resource limits apply to executions, local GPUs are managed by the NVIDIA GPU Operator and GPU time is accounted with the DCGM exporter.
High availability declared per service
In the Navin Manager, every service descriptor of the catalog must declare a high availability block, with its prerequisites and replicas. Installation checks refuse an HA deployment whose requirements are not met, and the dashboard of each service shows its HA status.
- HPA and KEDA for autoscaling (brick H05).
- NATS JetStream for event triggers (brick C10).
- PostgreSQL checkpoints and heartbeats for durable execution (bricks C02, C06 and C08).
- Quota service and GPU accounting (bricks H01 and H04).