The main design choice is that sessions share a host rather than each being a process. Agents are state in DB and the host handles the active work through a share event loop.
LLMs are getting better at coordination and teams (swarms) of agents are coming. Memory is getting expensive and my machine is always using swap. We cannot be wasting hundreds of MBs of RAM for an orchestrator (AKA the harness) of LLMs and tools.
More information on the architecture and measurement methodology is at https://divyanshgolyan.com/posts/hundred-agents/
I'd love feedback about this design boxing me into a corner and any large tradeoffs I'm not seeing!