AI Agents on GKE,
Powered by Gemini
Sympozium is a Kubernetes-native AI agent orchestration platform — forked and rebuilt from the ground up for Google Cloud. Vertex AI Gemini, Cloud Pub/Sub, Workload Identity, and self-hosted Gemma on TPU nodes. No non-Google dependencies.
Vertex AI Gemini
Three model tiers — Flash for speed, Pro for balance, Preview for power. Default is gemini-2.5-pro across all agent personas.
Hybrid Inference
Route between Vertex AI and self-hosted Gemma 3 models on GKE GPU (L4, A100) and TPU v5e nodes. Node-probe auto-discovery.
Cloud Pub/Sub Event Bus
NATS JetStream replaced with Cloud Pub/Sub for event-driven agent IPC. Topic auto-creation, trace context propagation, at-least-once delivery.
Workload Identity
Keyless authentication everywhere. No service account keys. GKE pods authenticate to GCP APIs via Workload Identity Federation.
Google Chat Channel
Cards v2, slash commands, threading, typing indicators, interactive retry buttons. Native integration with Google Workspace.
Observability
OpenTelemetry → Cloud Trace + Cloud Monitoring. Google Managed Prometheus + Grafana dashboards for existing Prometheus stacks.
Model Tier System
Every agent persona maps to a model tier. Route the right model to the right task.
High-throughput, low-latency tasks. Monitoring, log tailing, status checks.
General-purpose reasoning. Code review, incident response, architecture decisions.
Complex multi-step tasks. Security audits, cross-system analysis, planning.
Self-hosted on GKE GPU/TPU nodes. Air-gapped environments, cost optimization.