Hacker News new | ask | show | jobs
by tough 508 days ago
Yes, those are called multi-agent systems https://langchain-ai.github.io/langgraphjs/concepts/multi_ag...

there's also other approaches like swarms, etc

basically the same thing going on recursively with a singular -agent- having multiple -sub-skills- it can summon at will, the first prompt is an llm iterating over what tools to use given inputs, which can either reply with text, use a tool to get more context, or pass the message to another -agent-

so basically yeah most complex systems are multi-agentic in nature but maintain only one personality/tone in the frontend so it's less noticeable from a ux point of view.

Chatgpt for example, o1 is agent (does a thinking step)

also now models like r1 implementend it directly onto the streaming

there's also interesting stuff on continous thought research where the internal monlogue remains as tokens and not outputed, but still useful for the agents to -think step by step-

what is -an -agent- is a greqt question, simonwillson has compiled hundreds of answer people give along the way

1 comments

This sounds cool,

I didn't know about existing solutions in o1 and r1

I thought it was the same logic as input > thinking - researching - etc > output

Thanks for such a detailed answer and reference! I will check it