sympozium-ai/sympozium

Sympozium – Kubernetes‑native coordination layer for multi‑agent AI

What it is

• An open‑source control plane that runs on Kubernetes and lets you orchestrate many LLM‑powered agents the way K8s orchestrates containers.
• Agents are deployed as isolated workloads (pods, micro‑VMs called Celln). All policies –‑ what agents can see, which models they may use, which secrets they may touch –‑ are expressed as Custom Resource Definitions (CRDs).

Key ideas

Feature	Why it matters
Synthetic Membrane	Fine‑grained sharing rules (trust groups, visibility tags, field‑level gating) give teams of agents a shared situational awareness without leaking everything.
Ensembles	Bundle a set of agents, schedules, and memory into a Helm‑like package you can activate with a single CRD.
Shared Workflow Memory	SQLite‑based memory pool (with FTS5) that agents can read/write, with per‑agent access control and automatic decay.
Model Endpoints as CRDs	Declare a model (GGUF file, HuggingFace repo, etc.) once; Sympozium spins up an inference server (vLLM, llama.cpp, TGI…) and exposes an OpenAI‑compatible endpoint. Keys are mediated so the agent pods never see provider secrets.
Skill Sidecars	Each “skill” runs in its own least‑privilege sidecar container, garbage‑collected after the task finishes.
Celln fleet	On nodes with /dev/kvm the platform launches hardware‑isolated micro‑VMs (Celln) for long‑running agents, giving strong isolation and fast start‑up via warm pools.
Multi‑Channel adapters	Built‑in Deployments for Telegram, Slack, Discord, WhatsApp backed by NATS JetStream, turning chat messages into agent events.
Policy & Governance CRDs	SympoziumPolicy lets operators enforce allow/deny lists, network egress rules, image‑registry whitelists, and secret‑access controls on every agent run.
Observability	OpenTelemetry traces, Prometheus metrics, per‑run token‑usage cost estimates.
CLI / TUI / Web UI	sympozium command installs the stack, launches a terminal UI, or serves a browser dashboard for visual workflow canvases.

How you use it

1. Install – via Homebrew, a shell script, or Helm (CRDs first, then control‑plane). The installer checks for KVM‑enabled nodes and sets up the Celln fleet.
2. Define resources – create CRDs for models, agents, ensembles, policies, etc.
3. Run agents – agents can be one‑shot Jobs, long‑lived HTTP services, or external harnesses that talk to the platform through a secure adapter.
4. Observe & iterate – UI shows live canvas of agent pipelines; metrics let you monitor cost and performance.

What it does not do

• It does not host the actual LLM inference; it claims models and hands them to a separate capability layer (e.g., llmfit-dra, vLLM, SGLang).
• It does not replace the serving engine; token routing is left to the inference server you plug in.

Who might need it

• Teams building complex AI workflows where multiple specialized agents need to pass context, enforce security, and run on shared Kubernetes infrastructure.
• Enterprises that want audit‑ready, policy‑driven AI orchestration across on‑prem or cloud clusters.

Status

• Actively developed; APIs may change. CI badge and recent releases indicate a working codebase.

License

• MIT.

---

All details above are taken directly from the repository’s README.