Atmosphere/atmosphere ? reverse-engineered prompt

Build me a Java based real time framework for AI agents that can live inside a normal server app and stream responses to real clients. I want a simple way to define an agent, ideally with something like an annotation, and have it work over WebSocket, SSE, long polling, and gRPC without rewriting the endpoint each time.

It should feel ready for real use, not just a toy. Include reconnect support, durable sessions, run history, replay, memory, multi agent coordination, and human approval pauses that can resume later. I also want governance features like auth checks, scoped access, cost limits, PII redaction, and a way to kill runs. Make it portable so the same agent can be exposed to browsers and also through MCP, A2A, AG UI, and common chat channels.

Please include a CLI to create and run sample apps, plus a small JavaScript client for browser use. Add good docs and runnable examples, and look up current docs online if you need to.