I'm guessing in a way it's a lot like telling your agent to always use a subagent for toolcalls, and have it summarize the results >>1192241 some tool results are CRAZY big! like "ok let me investigate this 50MB log file to see what's going on" isn't in and of itself a problem, but it becomes a problem when the entire rest of the session has 49MB worth of noise occupying context space
like, it's a super advanced grep cause it doesn't need to do keywords, it tells a parallel version of itself "this is what I'm looking into, is this file relevant at all?" and it can do that for like, a whole github repo cause each file individually in all likelihood fits into context, even if together they're WAY too big to fit
Samu ／人◕ ‿‿ ◕人＼
uh huh sumarize the tool result ok
Samu ／人◕ ‿‿ ◕人＼
pretty fancy
S C
well, at least I THINK that's how it works, anyway I'm not learned enough to fully grasp it, but that seems to be the gist based on what it says in the git https://github.com/AtomicBot-ai/atomic-agent
S C
I really do think a lot of progress in what models can accomplish from here on out is going to be harness-side the actual models themselves are already fucking cracked, and largely constrained by how and what information is fed into the context, it seems
obviously progress is gonna made in the models themselves, too, but what's already out there is almost certainly very constrained
