Agents are services
Agents publish typed A2A contracts, health, versions, and capabilities into a governed registry.
All work01 02 03
Platform· Private platform build
AgentYard
A control plane for composing, deploying, and operating multi-agent systems
A self-hostable platform that treats agents as governed services: register them, compose systems visually across ten orchestration patterns, attach MCP tools at runtime, deploy to Kubernetes, and operate them through a full observability stack.
Role · Platform architecture and engineering
System anatomy
01Multi-Agent Systems
02A2A
03MCP
04Agent Observability
The problem
Enterprise agent teams repeatedly rebuild registries, orchestration, deployment, and observability. AgentYard unifies those concerns while avoiding an expensive LLM supervisor for deterministic patterns that do not need one.
What the system does
Capability with an operating model.
Governed, versioned registry for framework-agnostic A2A agents
ReactFlow studio with ten multi-agent composition patterns and portable YAML
Five purpose-built execution engines for chains, fan-out, loops, rules, and supervisors
Kubernetes, Docker, existing-endpoint, and EC2 deployment targets
Mission Control with traces, SSE logs, pod state, chaos testing, schedules, and DLQ replay
Runtime MCP tool attachment without rebuilding agent images
Architecture
How it fits together.
01
React 19 + TypeScript Studio and Mission Control experience
02
FastAPI services for Registry, Systems, Engine, Auth, and Gateway
03
PostgreSQL and Redis backed execution and state
04
Jaeger, Prometheus, Grafana, and Loki observability stack
05
Kind/Kubernetes, local Docker registry, Keycloak OIDC, and namespace RBAC
Key decisions
Use the cheapest correct engine
Sequential, parallel, fallback, and evaluation patterns run deterministically; an LLM supervisor is reserved for genuinely dynamic routing.
Operations are first-class
Tracing, logs, recovery, chaos testing, and deployment state live beside composition—not in a separate afterthought.