WEBVTT 00:00.400 --> 00:01.120 Hey there. 00:01.160 --> 00:01.920 Eden here. 00:01.920 --> 00:06.000 And in this video I want to talk about context engineering. 00:06.520 --> 00:13.640 Now, if you've been working with AI agents and probably coding agents like Cursor and Cloud Code, 00:13.640 --> 00:19.240 or maybe you even developed AI agents for your companies or for your own use, and you probably know 00:19.240 --> 00:26.840 that basically, it all boils down to a prompt being sent to an LLM and a lot of engineering around 00:26.840 --> 00:27.200 it. 00:27.440 --> 00:34.640 So there is some truth for calling applications like cursor and cloud code, just wrappers around llms. 00:35.040 --> 00:41.880 However, to build really good wrappers requires a lot of deep knowledge and a lot of engineering work 00:42.200 --> 00:47.640 because we know that those calls to those llms, they come with context. 00:47.640 --> 00:50.440 And this context comes from various sources. 00:50.680 --> 00:53.800 Context can come from the developer of the application. 00:53.800 --> 00:55.560 It can come from the user. 00:55.560 --> 01:01.640 It can come from the previous interaction of the user, from tool calls and other external data. 01:01.920 --> 01:10.210 The number of context sources is is increasing every day and sending the correct and relevant context 01:10.250 --> 01:11.210 to the LM. 01:11.690 --> 01:15.730 It's not quite as simple as we thought it was in the early days. 01:15.730 --> 01:20.810 We thought, hey, we have prompt engineering and we'll write some fancy prompts and this will fix all 01:20.810 --> 01:21.690 of the problem. 01:22.210 --> 01:25.930 However, the problem is this is that prompts are static. 01:25.930 --> 01:33.050 However, those pieces of context are extremely dynamic, and if they're extremely dynamic, then in 01:33.050 --> 01:39.290 order to construct the correct contents, we need to have a dynamic system as well. 01:39.290 --> 01:41.570 So it's not just a static prompt. 01:41.730 --> 01:45.930 And this is why we're entering the realm of context engineering. 01:45.930 --> 01:49.250 And it's the natural evolution of prompt engineering. 01:49.250 --> 01:52.610 But it's a much deeper concept. 01:53.090 --> 01:56.450 Now we all know the saying garbage in, garbage out. 01:56.450 --> 02:02.450 And this is a common reason why agentic systems don't perform as the way they should, because they're 02:02.450 --> 02:04.770 simply not provided with the right context. 02:05.210 --> 02:07.370 Llms cannot read our minds. 02:07.530 --> 02:10.410 We actually need to give them the right information. 02:10.610 --> 02:13.250 And by the way, it's not always information and data. 02:13.250 --> 02:18.650 Sometimes we need to give them the correct tools so they'll be able to fetch other information and take 02:18.650 --> 02:22.810 some actions and do some stuff for us, and then they'll achieve the task. 02:23.970 --> 02:26.370 So let's focus a bit on agents. 02:26.610 --> 02:29.610 And Llms are becoming better and better. 02:29.610 --> 02:30.890 And this is not new. 02:31.130 --> 02:33.650 They can reason very very well. 02:33.650 --> 02:36.730 And we have tool calling and we can build with it. 02:36.730 --> 02:43.650 Very cool functionality of AI agents which are running tools, invoking those tools, getting the output 02:43.650 --> 02:47.810 of those tools and then running in a loop until they finish a task. 02:48.290 --> 02:55.090 However, when it comes to a long running task and a complex task, we often accumulate the feedback 02:55.090 --> 03:00.970 from the tool calls, and this means that the context window is going to keep growing and growing, 03:01.170 --> 03:05.130 and to have lots of tokens being filled up with all the tool calls results. 03:05.410 --> 03:11.820 So this can cause a lot of problems, and it can of course exceed the size of the context window Though, 03:11.980 --> 03:15.420 and it can increase the cost, the latency. 03:15.660 --> 03:20.500 And by the end of it, it's eventually, if we're not going to do anything about it, it's going to 03:20.500 --> 03:24.380 degrade the agent performance, this degradation. 03:24.380 --> 03:25.580 There are many reasons. 03:25.580 --> 03:28.260 So for once we can have a context poisoning. 03:28.260 --> 03:36.140 So this is when one tool call or one call introduces a hallucination that makes it into the context. 03:36.140 --> 03:38.300 And it's starting now to degrade the system. 03:38.340 --> 03:40.420 Another thing that can happen is context. 03:40.420 --> 03:47.020 Confusion is that when we introduce some unnecessary context that is going to influence the response. 03:47.020 --> 03:50.100 So this is context which is not needed to the task. 03:50.580 --> 03:57.900 And of course there is the possibility of a context clash when parts of the context contradict each 03:57.940 --> 03:58.340 other. 03:59.300 --> 03:59.740 All right. 03:59.740 --> 04:02.340 So let me summarize this video in a couple of sentences. 04:02.340 --> 04:09.940 We discussed what is context engineering and in a very simple term is simply a way to give the LLM the 04:09.940 --> 04:11.060 correct context. 04:11.100 --> 04:11.780 We discussed. 04:11.780 --> 04:17.700 Where did this concept evolve from prompt engineering and why is it specifically important for agents? 04:17.700 --> 04:20.980 And in the next video we're going to discuss some techniques. 04:21.180 --> 04:25.620 And those techniques are going to give eventually the LM better context. 04:25.860 --> 04:30.380 Now some of those techniques are going to be on the application developer side. 04:30.380 --> 04:36.380 So for example, developers of applications like cloud code will implement those kinds of solution. 04:36.580 --> 04:40.500 However some of those techniques are on the user side. 04:40.500 --> 04:47.100 So we as users of cloud code has a lot of influence on the answer that we get and the context that we're 04:47.100 --> 04:49.100 going to supply the LM eventually. 04:49.700 --> 04:55.660 So this means that even non developers need to know context engineering and need to understand this 04:55.660 --> 05:01.700 principle if they want to get better responses and better answers from their AI systems and AI agents. 05:02.020 --> 05:08.220 Just to finish, an excellent example of context engineering techniques and how to implement them from 05:08.220 --> 05:13.980 both the developer side and both the user side is coding agents like cloud Code. 05:13.980 --> 05:15.940 And this is what we're going to be discussing. 05:15.940 --> 05:20.420 In the next video, I'll be discussing how to better engineer our context.