purpose is to significantly improve the tools’ capabilities;
the initial idea mostly applies to “open” LLMs, but could be expanded to save on costs or might even improve quality of output when using “closed” LLMs:
there are a number of ideas behind this concept (using more than 1 LLM), usually related to message chaining;
my idea is related to overcoming an LLM’s context, by using a LLM with a very large context window to take in the whole context, then extract the relevant info out of that and generate a very concise message to provide to the coder LLM(s);
this is mainly done whenever user sends a message, but could also be done internally in a loop, developing a task requiring many steps into individual steps if possible, each one then provided only the context it absolutely requires to complete the individual step.