WEBVTT 1 00:00:00.000 --> 00:00:02.400 So let's review the characteristics 2 00:00:02.400 --> 00:00:03.660 of Deep Agents. 3 00:00:03.660 --> 00:00:08.190 And I want to show you a concrete example how they manifest. 4 00:00:08.190 --> 00:00:11.580 And while Deep Agents is a general term, 5 00:00:11.580 --> 00:00:13.920 I think the ones who coined this term 6 00:00:13.920 --> 00:00:15.630 were the LangChain team. 7 00:00:15.630 --> 00:00:18.780 And they did an amazing job articulating this 8 00:00:18.780 --> 00:00:21.030 and coming up with this architecture 9 00:00:21.030 --> 00:00:25.140 after analyzing a bunch of Deep Agents implementations. 10 00:00:25.140 --> 00:00:27.510 All right, so let's go back to the architecture 11 00:00:27.510 --> 00:00:30.813 and let's go and talk about what makes a Deep Agent. 12 00:00:30.813 --> 00:00:33.720 And let's start with the planning tool. 13 00:00:33.720 --> 00:00:36.300 So every Deep Agent you are going to see, 14 00:00:36.300 --> 00:00:38.280 it's going to have a planning tool. 15 00:00:38.280 --> 00:00:41.490 And here we can see as an example, Claude Code. 16 00:00:41.490 --> 00:00:45.390 And here we have a bunch of tasks that are already done. 17 00:00:45.390 --> 00:00:48.120 We have one task that is in the making 18 00:00:48.120 --> 00:00:49.650 and is currently executing, 19 00:00:49.650 --> 00:00:53.490 and after it we have other tasks which needs to be executed. 20 00:00:53.490 --> 00:00:57.030 So what we're seeing here is not implicit planning 21 00:00:57.030 --> 00:00:59.040 through chain of thought reasoning 22 00:00:59.040 --> 00:01:01.560 we know from large language models. 23 00:01:01.560 --> 00:01:05.280 Deep Agents use explicit planning tools. 24 00:01:05.280 --> 00:01:07.740 And this is typically being implemented 25 00:01:07.740 --> 00:01:11.190 as a to-do list in a markdown format. 26 00:01:11.190 --> 00:01:15.780 And between execution, the agent is going to actively review 27 00:01:15.780 --> 00:01:17.580 and update this plan. 28 00:01:17.580 --> 00:01:19.980 So this plan is highly dynamic. 29 00:01:19.980 --> 00:01:22.440 The agent is going to mark tasks 30 00:01:22.440 --> 00:01:26.070 as pending, in progress or completed. 31 00:01:26.070 --> 00:01:29.310 And if we're going to have a task that failed, 32 00:01:29.310 --> 00:01:33.150 the agent is not going to blindly retry to do it 33 00:01:33.150 --> 00:01:36.450 like we know from the original React algorithm. 34 00:01:36.450 --> 00:01:40.890 So the planning tool is going to help us steer the agent. 35 00:01:40.890 --> 00:01:43.200 It's going to be continuously updated. 36 00:01:43.200 --> 00:01:46.770 And we as the users, we can also influence 37 00:01:46.770 --> 00:01:49.590 the task list and the to-do list. 38 00:01:49.590 --> 00:01:52.590 And in Claude Code specifically, 39 00:01:52.590 --> 00:01:54.870 the planning tool is internal, 40 00:01:54.870 --> 00:01:57.270 so we don't have direct access to it, 41 00:01:57.270 --> 00:01:59.370 however we can see it executing. 42 00:01:59.370 --> 00:02:00.870 And what you're seeing right now 43 00:02:00.870 --> 00:02:04.761 is a X post by Boris Cherny, which is the creator 44 00:02:04.761 --> 00:02:07.070 of Claude Code by the way. 45 00:02:07.070 --> 00:02:08.910 And here you can see the update to do, 46 00:02:08.910 --> 00:02:12.660 which is calling the planning tool to update the to-do list. 47 00:02:12.660 --> 00:02:15.630 And this to-do list is going to be continuously updated 48 00:02:15.630 --> 00:02:16.860 by the Deep Agent. 49 00:02:16.860 --> 00:02:19.740 And this is going to yield us better results 50 00:02:19.740 --> 00:02:22.680 and to actually accomplish those complex tasks. 51 00:02:22.680 --> 00:02:25.590 And when thinking about it, this is pretty intuitive 52 00:02:25.590 --> 00:02:28.320 because we as humans, if we have a complex task, 53 00:02:28.320 --> 00:02:30.720 we usually want to break it down. 54 00:02:30.720 --> 00:02:34.230 We want to keep track of what we have accomplished so far. 55 00:02:34.230 --> 00:02:36.360 So this also give us a dopamine rush, 56 00:02:36.360 --> 00:02:39.390 you know, to check a task that is completed. 57 00:02:39.390 --> 00:02:42.300 And it really shows the progress of our project 58 00:02:42.300 --> 00:02:45.240 and something complex that we are trying to achieve. 59 00:02:45.240 --> 00:02:47.913 So the idea here is actually very intuitive.