3 ms·
One of the most valuable techniques for building useful LLM systems right now is actually the opposite of that. Context is limited in length and too much stuff
by simonw 1y ago
One of the most valuable techniques for building useful LLM systems right now is actually the opposite of that.
Context is limited in length and too much stuff in the context can lead to confusion and poor results - the solution to that is "sub-agents", where a coordinating LLM prepares a smaller context and task for another LLM and effectively treats it as a tool call.
The best explanation of that pattern right now is this from Anthropic: https://www.anthropic.com/engineering/built-multi-agent-research-system https://www.anthropic.com/engineering/built-multi-agent-rese...