sympozium-ai/sympozium

What it solves

Sympozium provides a Kubernetes-native coordination layer for multi-agent AI systems. It addresses the challenge of managing multiple agents that need to share context, hand off tasks, and maintain shared situational awareness, treating agents as Pods and policies as Custom Resource Definitions (CRDs).

How it works

Sympozium acts as a control plane that manages the execution and interaction of AI agents. It uses Kubernetes primitives to handle agent deployment, where every execution is a Job and every agent is a Pod. It integrates with various inference engines (like vLLM or llama.cpp) and AI providers (OpenAI, Anthropic, etc.) without vendor lock-in. Key mechanisms include:

• Synthetic Membrane: Controls what information agents share via trust groups and visibility tags.
• Ensembles: Bundles of agent teams that can be activated as a single pack.
• Shared Workflow Memory: A pack-level SQLite memory pool for cross-persona knowledge sharing.
• Skill Sidecars: Runs agent skills in isolated sidecars with ephemeral RBAC.
• Agent Sandbox: Provides kernel-level isolation (via gVisor or Kata) for secure agent execution.

Who it’s for

It is designed for developers and platform engineers building complex multi-agent AI systems that require robust orchestration, security, and scalability on Kubernetes.

Highlights

• K8s-native architecture: Uses CRDs and the stock Kubernetes scheduler for agent and model placement.
• Multi-channel support: Native integration with Telegram, Slack, Discord, and WhatsApp via NATS JetStream.
• MCP Server integration: Supports the Model Context Protocol for external tool discovery and filtering.
• Comprehensive governance: Cluster-wide SympoziumPolicy for tool gating, network egress, and image-registry allowlists.
• Observability: Built-in OpenTelemetry traces, Prometheus metrics, and per-run token cost estimation.