1
00:00:00,180 --> 00:00:06,160
And I started going down, my prediction was there's a major amount of context engineering

2
00:00:06,160 --> 00:00:09,940
that's happening. You don't wanna send the entire history

3
00:00:11,260 --> 00:00:16,780
to the LLM every time. But there's this really awesome optimization, and thank you for

4
00:00:16,780 --> 00:00:21,560
open sourcing it, because we can kind of see and take inspiration from this, is that

5
00:00:21,560 --> 00:00:28,120
every time the model requests a tool call, the harness makes a tool call, whatever,

6
00:00:28,860 --> 00:00:35,100
sometimes you're getting Git repo and you're publishing code inside that prompt. And

7
00:00:35,100 --> 00:00:39,620
what happens is over time, especially as your history gets longer and longer,

8
00:00:41,180 --> 00:00:47,000
that context gets longer and longer. And all you have is a bunch of lines of code that

9
00:00:47,000 --> 00:00:51,560
you referenced 20 minutes ago that have nothing to do with the task that you're working

10
00:00:51,560 --> 00:00:57,180
on, but K -clot code is sending that every time. What Strands is doing is, and it's all

11
00:00:57,180 --> 00:01:02,260
configurable, which was the other really nice thing. It just truncates tool context to

12
00:01:02,260 --> 00:01:07,680
a, each tool output is like by default to 1 ,000 characters, I think.

13
00:01:10,460 --> 00:01:15,060
And that means that if the model actually needs the code that was referenced there, it

14
00:01:15,060 --> 00:01:19,180
can ask for it. You can go get it again. But if it doesn't need it, don't bother the

15
00:01:19,180 --> 00:01:24,660
model sending it. And this is actually in terms of just like, it's a huge service to

16
00:01:24,660 --> 00:01:29,800
GPU capacity around the world to do this kind of stuff instead of just sending the entire

17
00:01:29,800 --> 00:01:30,220
thing.