Aj2280/advanced-multi-agent — reverse-engineered prompt

Build me a local advanced multi agent assistant that feels production ready. I want one main orchestrator that can spin up a researcher, coder, and analyst to work together on bigger tasks, either by debating or by passing work through a pipeline. Give me both a simple chat style web app and a command line way to run it.

The agents should keep memory in layers so they can hold short context, remember the current session, and also recall older useful information later. Let the researcher search the live web when needed, and let the coder run Python safely in an isolated sandbox with time limits. If anything is sensitive or marked critical, pause and ask me for approval before continuing.

I also want it to be reliable and efficient, with provider fallback, caching for repeated questions, async communication between agents, and monitoring so I can inspect traces, metrics, and dashboards. Make it easy to start locally with Docker and simple config files for agent behavior and model routing. Look up current docs online if you need to.