I think multi-agent (eg _different_ underlying LLMs) everything is really the future. Code produced via multi-agent workflows and reviews seems noticeably better. I've been experimenting with a multi-agent message board recently: https://github.com/pjlsergeant/dogpark
I've built agents that call different LLMs and keep separated memories. Remember, agents are just long-running workflows with some nodes calling LLMs and that sometimes can be started as tool from other "agent".
There are times when I wonder if couldn't just draw then in a BPMN designer that allowed me to write custom code for nodes. Is BPMN still a thing?
Experimented with multi-llm analysis for problem solving over summer, combined with multi-agent approaches it can tease out interesting angles to problems that I never considered. Expensive but use only for my high value problems.
Comments
I think multi-agent (eg _different_ underlying LLMs) everything is really the future. Code produced via multi-agent workflows and reviews seems noticeably better. I've been experimenting with a multi-agent message board recently: https://github.com/pjlsergeant/dogpark
I've built agents that call different LLMs and keep separated memories. Remember, agents are just long-running workflows with some nodes calling LLMs and that sometimes can be started as tool from other "agent".
There are times when I wonder if couldn't just draw then in a BPMN designer that allowed me to write custom code for nodes. Is BPMN still a thing?
Experimented with multi-llm analysis for problem solving over summer, combined with multi-agent approaches it can tease out interesting angles to problems that I never considered. Expensive but use only for my high value problems.
https://github.com/monkeydust/rightmind
There are good arguments for subagents/multi-agents. But the added overhead is usually massive.