WEBVTT 1 00:00:00.690 --> 00:00:02.460 Before we talk about deep agents, 2 00:00:02.460 --> 00:00:06.030 we need first to understand a bit of taxonomy 3 00:00:06.030 --> 00:00:09.300 and to talk about what kind of agents do we have 4 00:00:09.300 --> 00:00:12.000 currently in production in the industry. 5 00:00:12.000 --> 00:00:15.930 So let me simplify this and give you how I view things. 6 00:00:15.930 --> 00:00:18.630 So we have the domain of agents 7 00:00:18.630 --> 00:00:21.540 and here are going to be all types of agents 8 00:00:21.540 --> 00:00:23.160 we have now in the industry. 9 00:00:23.160 --> 00:00:26.970 Now they can be agents, they can be agentic applications 10 00:00:26.970 --> 00:00:30.510 like for example this hybrid RAG architecture, 11 00:00:30.510 --> 00:00:34.620 where we have an LLM which is deciding which step to make, 12 00:00:34.620 --> 00:00:39.210 whether to make a search, whether to rephrase the query 13 00:00:39.210 --> 00:00:40.530 before the retrieval. 14 00:00:40.530 --> 00:00:43.770 And we have an LLM making a decision what's going 15 00:00:43.770 --> 00:00:46.080 to be the next step to be executed. 16 00:00:46.080 --> 00:00:49.380 Now under this umbrella of agents, 17 00:00:49.380 --> 00:00:54.380 we have here the ReAct agent where we have an LLM, 18 00:00:54.480 --> 00:00:58.050 which is deciding whether to use a tool or not. 19 00:00:58.050 --> 00:01:01.860 We execute this tool, then we either decide that we want 20 00:01:01.860 --> 00:01:03.960 to return a response to the user if we have enough 21 00:01:03.960 --> 00:01:06.540 information and if not, this agent continues 22 00:01:06.540 --> 00:01:09.960 an execute tool gets the result, the observation 23 00:01:09.960 --> 00:01:14.130 and simply goes on and on and on until it gets an answer. 24 00:01:14.130 --> 00:01:17.280 And by the way, this is what started this all, 25 00:01:17.280 --> 00:01:19.140 this all agent paradigm started 26 00:01:19.140 --> 00:01:22.620 from this ReAct algorithm, this ReAct loop here. 27 00:01:22.620 --> 00:01:24.840 Now when I showed you the taxonomy 28 00:01:24.840 --> 00:01:27.960 and show you how I view the words of agents, 29 00:01:27.960 --> 00:01:32.130 I refer to the ReAct agent as a shallow agent. 30 00:01:32.130 --> 00:01:35.640 Now the reason why I refer to it as a shallow agent 31 00:01:35.640 --> 00:01:39.870 is because this agent doesn't have the capabilities 32 00:01:39.870 --> 00:01:41.730 of going really, really deep. 33 00:01:41.730 --> 00:01:44.100 So for example, to perform a very deep 34 00:01:44.100 --> 00:01:47.640 and thorough research, even though it has a Search tool 35 00:01:47.640 --> 00:01:49.680 and a Wiki tool and a Tavily tool, 36 00:01:49.680 --> 00:01:53.700 if I want to make a decent research, 37 00:01:53.700 --> 00:01:55.980 this agent is not going to be enough. 38 00:01:55.980 --> 00:01:59.190 And the reason for it is from this agent design 39 00:01:59.190 --> 00:02:02.760 and from the reality of modern LLMs. 40 00:02:02.760 --> 00:02:06.120 So this agent is based on function calling. 41 00:02:06.120 --> 00:02:08.970 And in this decision arrow here we have here, 42 00:02:08.970 --> 00:02:11.580 there is a function call that the LLM makes 43 00:02:11.580 --> 00:02:13.860 in order to choose the tool to execute. 44 00:02:13.860 --> 00:02:15.180 Now don't get me wrong, 45 00:02:15.180 --> 00:02:17.730 there is no problem with function calling. 46 00:02:17.730 --> 00:02:19.890 Actually there is a bit of problem with function calling. 47 00:02:19.890 --> 00:02:22.170 I discussed it when I talk about code mode. 48 00:02:22.170 --> 00:02:24.870 But the inherent limitation of this agent, 49 00:02:24.870 --> 00:02:26.160 what makes it shallow 50 00:02:26.160 --> 00:02:30.090 and not being able to perform deep tasks is that 51 00:02:30.090 --> 00:02:32.910 it has a very limited context window. 52 00:02:32.910 --> 00:02:36.300 Now we know LLMs have context window. This is not new. 53 00:02:36.300 --> 00:02:39.720 But because of this architecture, every time we're going 54 00:02:39.720 --> 00:02:42.720 to be making a decision then to execute the tool, 55 00:02:42.720 --> 00:02:45.060 then to plug the result into the LLM, 56 00:02:45.060 --> 00:02:48.660 we are going to be bloating the context. 57 00:02:48.660 --> 00:02:50.880 And if we're going to have one iteration, 58 00:02:50.880 --> 00:02:52.710 it's going to be maybe reasonable. 59 00:02:52.710 --> 00:02:55.200 But as we have more iterations, 60 00:02:55.200 --> 00:02:58.320 this context window keeps growing and growing and growing 61 00:02:58.320 --> 00:03:01.290 and this is going to be leading to context rot. 62 00:03:01.290 --> 00:03:04.830 So we can have context confusion, context contradiction, 63 00:03:04.830 --> 00:03:08.550 context pollution, and a lot of things that eventually 64 00:03:08.550 --> 00:03:11.490 are going to degrade the LLM's performance 65 00:03:11.490 --> 00:03:16.490 and are going to get our agents to go off rails. 66 00:03:16.500 --> 00:03:18.870 Just to emphasize, I'm talking here about the use case 67 00:03:18.870 --> 00:03:23.310 where we have a complex task, a deep task where we need 68 00:03:23.310 --> 00:03:27.150 to continue and continue and continue and gather information 69 00:03:27.150 --> 00:03:29.040 and deduce and process. 70 00:03:29.040 --> 00:03:31.830 And we have something which is long running, 71 00:03:31.830 --> 00:03:33.660 not simply book me a flight. 72 00:03:33.660 --> 00:03:35.580 We have something much more complex like 73 00:03:35.580 --> 00:03:36.660 to implement a feature 74 00:03:36.660 --> 00:03:39.930 or to implement a deep research on a topic. 75 00:03:39.930 --> 00:03:43.110 So up until now I just talk about the degradation 76 00:03:43.110 --> 00:03:45.170 of the performance of the LLM, 77 00:03:45.170 --> 00:03:47.490 of the quality of the results. 78 00:03:47.490 --> 00:03:49.380 Of course it's going to be costing us more money 79 00:03:49.380 --> 00:03:51.960 because it's more tokens for each call we make. 80 00:03:51.960 --> 00:03:55.050 So the LLM call is going to be heavier. 81 00:03:55.050 --> 00:03:56.850 So it's going to contain more tokens, 82 00:03:56.850 --> 00:03:59.670 it's going to cost us more, it's going to be also slower. 83 00:03:59.670 --> 00:04:03.720 And this agent, this ReAct agent architecture, 84 00:04:03.720 --> 00:04:05.190 while it is amazing 85 00:04:05.190 --> 00:04:07.980 and it's the basis for basically everything, 86 00:04:07.980 --> 00:04:12.980 it's not really usable for complex tasks, those deep tasks. 87 00:04:12.990 --> 00:04:15.540 It is actually very good for shallow tasks 88 00:04:15.540 --> 00:04:17.850 that don't require a lot of iterations. 89 00:04:17.850 --> 00:04:19.770 And we actually have a lot of those kinds 90 00:04:19.770 --> 00:04:21.120 of agents in production. 91 00:04:21.120 --> 00:04:22.770 I've worked with many customers 92 00:04:22.770 --> 00:04:25.410 and a lot of them have this kind of architecture. 93 00:04:25.410 --> 00:04:28.590 And for a lot of use cases, it is actually enough 94 00:04:28.590 --> 00:04:30.810 and we don't really need more than that. 95 00:04:30.810 --> 00:04:33.270 Let's go and look at the world of agents again 96 00:04:33.270 --> 00:04:36.990 and we explain this shallow and ReAct part here. 97 00:04:36.990 --> 00:04:40.050 So now let's start to talk about deep agents. 98 00:04:40.050 --> 00:04:43.710 And deep agents are agents that are able 99 00:04:43.710 --> 00:04:46.830 to perform those kinds of long horizon tasks. 100 00:04:46.830 --> 00:04:50.460 Those tasks that are complex that require a lot 101 00:04:50.460 --> 00:04:53.370 of iterations, a lot of processing. 102 00:04:53.370 --> 00:04:55.020 Those agents are long running. 103 00:04:55.020 --> 00:04:59.220 They can run for minutes, for hours, even for days. 104 00:04:59.220 --> 00:05:01.170 They can get even user input. 105 00:05:01.170 --> 00:05:04.320 They can stop their execution, they can then resume it 106 00:05:04.320 --> 00:05:06.210 after they receive the user input. 107 00:05:06.210 --> 00:05:09.240 And those are what are called deep agents. 108 00:05:09.240 --> 00:05:12.150 So we have here deep research agents 109 00:05:12.150 --> 00:05:14.370 and here for example, we have Perplexity 110 00:05:14.370 --> 00:05:17.280 and we have here a deep research feature. 111 00:05:17.280 --> 00:05:20.250 And this is going to trigger a deep research agent run. 112 00:05:20.250 --> 00:05:23.310 Now deep research agents are very, very common today. 113 00:05:23.310 --> 00:05:26.400 So almost every other vendor have kind of support 114 00:05:26.400 --> 00:05:27.840 for research agents. 115 00:05:27.840 --> 00:05:31.350 Here we can see in Claude Code the research option. 116 00:05:31.350 --> 00:05:35.550 And here we can see in ChatGPT also a research option. 117 00:05:35.550 --> 00:05:38.640 And all of those are going to trigger deep agents 118 00:05:38.640 --> 00:05:40.470 that are going to perform research. 119 00:05:40.470 --> 00:05:43.260 And every one of them is implemented a bit differently 120 00:05:43.260 --> 00:05:44.910 by the vendors. 121 00:05:44.910 --> 00:05:49.020 We have also open source deep research agents, 122 00:05:49.020 --> 00:05:51.660 for example GPT Researcher, which at the time 123 00:05:51.660 --> 00:05:53.010 of recording this video 124 00:05:53.010 --> 00:05:55.740 is one of the most popular open source projects. 125 00:05:55.740 --> 00:05:59.130 And I actually review it in depth in my LangGraph course. 126 00:05:59.130 --> 00:06:01.740 Alright, let's go back to the taxonomy here. 127 00:06:01.740 --> 00:06:05.280 So we talked about deep research agents as an example, 128 00:06:05.280 --> 00:06:09.900 and we also have coding agents like Claude Code, like Devin, 129 00:06:09.900 --> 00:06:14.220 like Cursor, like Gemini CLI, and the list goes on and on 130 00:06:14.220 --> 00:06:18.180 because coding agents are actually one of the best examples 131 00:06:18.180 --> 00:06:21.060 of how deep agents can be very, very effective 132 00:06:21.060 --> 00:06:22.770 and useful for the industry 133 00:06:22.770 --> 00:06:26.250 and actually used by millions of developers. 134 00:06:26.250 --> 00:06:28.620 And here I'm going to be focusing on Claude Code 135 00:06:28.620 --> 00:06:32.430 because it is currently the leading coding agent. 136 00:06:32.430 --> 00:06:36.690 Now Claude Code is extremely good, it's very, 137 00:06:36.690 --> 00:06:40.830 very useful and it can actually be used for a wide variety 138 00:06:40.830 --> 00:06:43.050 of tasks which are long-running. 139 00:06:43.050 --> 00:06:47.130 They need to perform deep, they need scale. 140 00:06:47.130 --> 00:06:50.010 And this is the top coding agent 141 00:06:50.010 --> 00:06:52.020 at least when making this video. 142 00:06:52.020 --> 00:06:54.330 So we can go and ask Claude 143 00:06:54.330 --> 00:06:57.900 or any other coding agent to implement us an application 144 00:06:57.900 --> 00:07:00.690 or a feature in the application and those agents can go 145 00:07:00.690 --> 00:07:02.730 and run and actually accomplish it. 146 00:07:02.730 --> 00:07:06.630 They can go and not only code the required code 147 00:07:06.630 --> 00:07:09.930 but also to run tests, to test the application, 148 00:07:09.930 --> 00:07:12.360 to open up a browser, to take screenshots 149 00:07:12.360 --> 00:07:14.580 and to do the very same things 150 00:07:14.580 --> 00:07:17.040 that a normal software engineer would do. 151 00:07:17.040 --> 00:07:22.040 So those coding agents are going to be a subset 152 00:07:22.170 --> 00:07:25.560 of deep agents which are tailor-made for coding. 153 00:07:25.560 --> 00:07:27.480 So this is their specialty. 154 00:07:27.480 --> 00:07:30.570 And as a side note here, deep agents are 155 00:07:30.570 --> 00:07:33.060 what today drive innovation 156 00:07:33.060 --> 00:07:36.510 and they are the top tier technology we have right now. 157 00:07:36.510 --> 00:07:39.720 And it's actually very interesting to see the fact 158 00:07:39.720 --> 00:07:42.600 that LLMs now are becoming better and better, 159 00:07:42.600 --> 00:07:46.140 but currently their improvements are pretty gradual. 160 00:07:46.140 --> 00:07:49.920 So we don't have this huge leap of reasoning 161 00:07:49.920 --> 00:07:51.780 or other capabilities 162 00:07:51.780 --> 00:07:54.780 that we see from each new model that comes out. 163 00:07:54.780 --> 00:07:56.550 The quality is getting better. 164 00:07:56.550 --> 00:08:00.210 Yes, the reasoning is better, the capabilities are better, 165 00:08:00.210 --> 00:08:01.530 but it grows gradually. 166 00:08:01.530 --> 00:08:02.940 It doesn't grow exponentially. 167 00:08:02.940 --> 00:08:04.980 However, we are now at the point 168 00:08:04.980 --> 00:08:09.480 that abstraction on abstractions of agents, on agents 169 00:08:09.480 --> 00:08:11.670 in the application layer, if we know how 170 00:08:11.670 --> 00:08:13.710 to use those LLMs correctly, 171 00:08:13.710 --> 00:08:15.870 we can achieve some very impressive 172 00:08:15.870 --> 00:08:19.590 and very capable machines that can automate 173 00:08:19.590 --> 00:08:23.070 a lot of human work, which requires a lot of reasoning. 174 00:08:23.070 --> 00:08:26.340 Things that we once thought that are impossible 175 00:08:26.340 --> 00:08:29.160 because if I were to tell you five years ago, for example, 176 00:08:29.160 --> 00:08:32.340 that we are going to have a technology that is able 177 00:08:32.340 --> 00:08:37.340 to create a working pretty beautiful application from zero 178 00:08:37.470 --> 00:08:40.290 with AI with no human intervention, 179 00:08:40.290 --> 00:08:42.360 you would probably say that I'm crazy. 180 00:08:42.360 --> 00:08:44.760 And the point I'm trying to make here is 181 00:08:44.760 --> 00:08:48.870 that the application layer today, the application layer 182 00:08:48.870 --> 00:08:51.330 that implements those kinds of deep agents, 183 00:08:51.330 --> 00:08:55.710 this is what's driving now the technology, not the LLMs 184 00:08:55.710 --> 00:08:57.060 that are getting better, 185 00:08:57.060 --> 00:09:01.260 but the application layer that we build as developers on top 186 00:09:01.260 --> 00:09:03.840 of them, how to harness those LLMs. 187 00:09:03.840 --> 00:09:06.810 And today, this is what pushes the boundaries 188 00:09:06.810 --> 00:09:09.330 and pounds on the innovation. 189 00:09:09.330 --> 00:09:12.270 This type of work, the application layer work, 190 00:09:12.270 --> 00:09:15.150 this deep agents implementation. 191 00:09:15.150 --> 00:09:18.450 And let's go back to the discussion on deep agents 192 00:09:18.450 --> 00:09:21.480 and what makes an agent a deep agent. 193 00:09:21.480 --> 00:09:24.240 So there isn't really a clear definition, 194 00:09:24.240 --> 00:09:28.410 but an agent that is able to perform a complex task, 195 00:09:28.410 --> 00:09:30.180 in my opinion is a deep agent. 196 00:09:30.180 --> 00:09:32.760 And in order for an agent to go 197 00:09:32.760 --> 00:09:35.580 and perform these kinds of research tasks 198 00:09:35.580 --> 00:09:38.610 or coding tasks which are complex 199 00:09:38.610 --> 00:09:40.980 and require long running, 200 00:09:40.980 --> 00:09:44.160 we need to have certain capabilities of those agents 201 00:09:44.160 --> 00:09:47.880 that are going to make sure that we get a quality result. 202 00:09:47.880 --> 00:09:51.300 Now it all boils down to context engineering 203 00:09:51.300 --> 00:09:55.470 and smart context management, which is going to allow us 204 00:09:55.470 --> 00:09:58.410 to solve this problem of the context bloat 205 00:09:58.410 --> 00:10:00.330 and context accumulation. 206 00:10:00.330 --> 00:10:04.350 And it's going to help us scale those agents 207 00:10:04.350 --> 00:10:07.320 to perform a lot of things concurrently. 208 00:10:07.320 --> 00:10:08.670 And to achieve this 209 00:10:08.670 --> 00:10:10.950 and to achieve this capability of performing 210 00:10:10.950 --> 00:10:13.530 and executing on long horizon tasks, 211 00:10:13.530 --> 00:10:17.160 most deep agents implement the following. 212 00:10:17.160 --> 00:10:21.480 First, they have a planning tool which is going to plan 213 00:10:21.480 --> 00:10:24.030 what the deep agent is going to do. 214 00:10:24.030 --> 00:10:27.270 They also have a subagents capability. 215 00:10:27.270 --> 00:10:30.660 So this means they can spawn specialized workers 216 00:10:30.660 --> 00:10:33.810 that are going to work in isolated context 217 00:10:33.810 --> 00:10:36.540 and going to help the deep agent scale 218 00:10:36.540 --> 00:10:39.300 and perform a lot of tasks concurrently 219 00:10:39.300 --> 00:10:41.550 without bloating the context. 220 00:10:41.550 --> 00:10:43.680 Thirdly, they have a file system, 221 00:10:43.680 --> 00:10:46.740 where they can write intermediate results 222 00:10:46.740 --> 00:10:49.530 and shared states between those agents 223 00:10:49.530 --> 00:10:52.170 and everything is not going to blow to the context 224 00:10:52.170 --> 00:10:54.210 because it's going to be written to the disc. 225 00:10:54.210 --> 00:10:58.650 And thirdly, they have a monstrous system prompt. 226 00:10:58.650 --> 00:11:01.980 If you're going to pick up today one deep agent, 227 00:11:01.980 --> 00:11:05.520 it's going to probably implement those kinds of four ideas. 228 00:11:05.520 --> 00:11:07.680 And in the next video, I'll be diving 229 00:11:07.680 --> 00:11:09.060 and explaining those ideas, 230 00:11:09.060 --> 00:11:11.110 including show you implementations of it.