Hi team, I would love to have your views on github.com/OpenHands/OpenHands/issues/16315 Specially the part of a structured knowledge management system that can be shared by multiple agents. Is there a small part of it I can start with that can show some immediate ROI that would be super helpful. Thanks in advance

Hey, thank you for the proposal! I left a comment on some details

We recently added the persistent memory in agent canvas, so we have that as well. You can always add more stuff like databases, but not sure if that is necessary for most use cases unless you get to really concurrent multiagent setups. It would be great if you could pressure test and see how much you can build on the existing parts. Do you have a benchmark or datasets we can test the performance against to see how well we can do. There was a recent Kaggle competition neurogolf where structured management across multiagents works really well. I played around with using openhands to implement it here: github.com/rajshah4/openhands-agent-research-lab and rajshah4.github.io/openhands-agent-research-lab (I also have other resources on multiagent if you are curious about orchestrating that in openhands)

Hi @calm-artist-7206 Thanks. I agree most of it is already in openhands or other coding agents like codex. I'll detail out my use case and maybe you can guide me on that. I have an application that runs certain validation rules on my data pipeline before ingesting. I want to design and test candidate rules on my already existing data. Agent 1 - Comes up with candidate rules Agent 2 - Runs in a sandbox environment with local LLM to test the rules and summarise breaks Agent 3 - Codes approved rules into the application I want a seamless communication between these agents running in my local and sandbox. I also want it to controlled as the data in my sandbox is sensitive and should not be exposed to external LLMs

Yea, I would think if you are keeping rules, you can use existing text based memory with md files to help keep track of that information. The OpenHands SDK has multiple approaches orchestrating those agents depending how long lasting they are. Happy to help as you set that up. (You could also have these agents running with our automation server, so whenever there is a new rule, it triggers the testing with an eveant based automation, and once the testing is over it triggers the approval process)