atmosphere

Why we included this project

JVM teams whose AI agents have to run like ordinary production services, not notebook demos, will find the missing layer here. A plain annotated agent becomes a deep one by default: memory, a plan, a virtual filesystem, and sub-agent delegation are all on, and the output streams to clients over WebSocket, SSE, long-polling, or gRPC through one broadcaster you can filter and gate. MCP, A2A, and AG-UI are spoken out of the box, and the runtime stays portable across frameworks: behind a single runtime SPI you can move between Spring AI, LangChain4j, Google ADK, and others without rewriting the endpoint. Governance is built in rather than bolted on later, covering policy admission, human approval, cost ceilings, and PII rewriting before each tool call. If your agent needs durable sessions, reconnect handling, and per-tenant observability inside an existing JVM stack, this is worth studying before you hand-roll the plumbing.