Your agents are passing sticky notes. Give them a control plane.

What it does Sympozium is a Kubernetes operator that treats AI agents as native cluster workloads: every agent is a Pod, every policy is a CRD, and every execution is a Job. It replaces ad-hoc message-passing with a shared coordination layer—SQLite memory pools with per-persona access control, structured workflow handoffs, and circuit breakers—so agent teams maintain collective situational awareness. The stack includes a TUI and web dashboard, supports local GGUF inference without API keys, and can expose agents through Slack, Telegram, Discord, or WhatsApp via NATS JetStream.

The interesting bit The project frames the central challenge as the “sticky-note problem”: agents that pass serialized messages remain blind to each other’s state. The fix is a “synthetic membrane”—a trust and visibility layer inspired by Kubernetes’ own shared control plane—that selectively gates what each agent can see and share, backed by a published research paper. Agent teams can also be packaged into Helm-like “Ensembles” that the controller instantiates automatically.

Key highlights

• Skill sidecars run with ephemeral least-privilege RBAC and are garbage-collected on completion
• Persistent memory uses SQLite + FTS5 on a PersistentVolume, surviving across ephemeral Pod runs
• Local model inference declares GGUF weights as CRDs and deploys llama-server with OpenAI-compatible endpoints
• Multi-channel support treats each chat platform as a dedicated Deployment backed by NATS JetStream
• Optional kernel-level sandboxing via gVisor or Kata Containers with warm pools for fast starts

Caveats

• Explicitly pre-stable: the README warns that APIs will change and things will break
• The README displays an MIT license badge but states “Apache License 2.0” at the bottom, creating a licensing ambiguity

Verdict Built by the creator of k8sgpt, this is a natural fit for platform teams already living in Kubernetes who want to orchestrate agent fleets like any other workload; pure application developers or those seeking a managed, turn-key service should probably wait for the paint to dry.