here is my extremely naive take:
I think that simple binary checks on the messages will be hard to make work (even if you are 100x cheaper, if you do a linear amount of work it will add up)
but if we designed the history hierarchically (imagine each task has a nested set of labels of the current task), I think doing a tree search on context could be 🔥
when context becomes a log(n) search, all of a sudden a new explosion of possibilities comes up (e.g. why not include your history in every coding agent context? why not include the context of coding agents working in parallel? of subagents?)
here is my extremely naive take:
I think that simple binary checks on the messages will be hard to make work (even if you are 100x cheaper, if you do a linear amount of work it will add up)
but if we designed the history hierarchically (imagine each task has a nested set of labels of the current task), I think doing a tree search on context could be 🔥
when context becomes a log(n) search, all of a sudden a new explosion of possibilities comes up (e.g. why not include your history in every coding agent context? why not include the context of coding agents working in parallel? of subagents?)