WEBVTT 00:00.160 --> 00:05.800 Well, it was great to have seen our essential use case in action and with its own user interface. 00:05.920 --> 00:07.240 So what's next? 00:07.400 --> 00:09.360 Well, next it's over to you. 00:09.600 --> 00:14.280 The thing we want to do is make life harder for our deep research agent. 00:14.280 --> 00:16.240 We want to put it to work. 00:16.560 --> 00:19.200 So we want to really take this to the next level. 00:19.200 --> 00:25.760 It's still reasonably surface level in terms of the substantive ness of these, this deep research reporting. 00:26.040 --> 00:31.480 And it's something that is giving you an easy starting point to turn this into something which is a 00:31.480 --> 00:33.120 fully baked product. 00:33.480 --> 00:39.040 And here are some of the things that I am asking you to do and give this a shot. 00:39.080 --> 00:40.640 Do do your best. 00:40.880 --> 00:45.280 So first of all, we can learn a lesson from the way that OpenAI does it themselves. 00:45.280 --> 00:50.840 One of the tricks that you'll see that it does is that when you first ask for a deep research challenge, 00:50.960 --> 00:55.000 the first thing that happens is it comes back and asks for some questions. 00:55.000 --> 00:57.040 And so it would be great for us to build that. 00:57.040 --> 00:57.720 You can imagine. 00:57.720 --> 00:59.080 That's very simple. 00:59.120 --> 01:00.080 Agent step. 01:00.080 --> 01:02.290 The first step is to ask, ask. 01:02.330 --> 01:10.370 Come up with three clarifying questions associated with this query, and then make sure that you incorporate 01:10.370 --> 01:14.010 those clarifying questions in what searches are done. 01:14.050 --> 01:15.570 Take them into account. 01:15.890 --> 01:22.810 But then the the hardest part of this is that you really want to turn that manager that we have from 01:22.810 --> 01:28.930 just being a Python script and a series of function calls into being something that's truly agentic. 01:29.210 --> 01:35.050 Take some hints from what we did earlier in the week when we were taking agents as tools, and we were 01:35.050 --> 01:41.850 doing handoffs and build some of that into the deep research so that it has some more autonomy to decide 01:41.850 --> 01:44.050 whether it wants to do more searches or not. 01:44.090 --> 01:47.810 You might want to put a cap on that, but you put some controls in there. 01:47.810 --> 01:54.330 But but you want to try and give it some ability to go off and explore and also potentially to refine 01:54.330 --> 01:56.730 the queries based on what it's learned so far. 01:56.930 --> 02:03.110 You could be thinking about adding in the kind of Evaluator optimizer design patterns, so that there's 02:03.190 --> 02:08.150 another agent that gets to check the work of the deep researcher as well. 02:08.430 --> 02:14.710 So all of these patterns that you can add, but most importantly, the overall autonomous manager to 02:14.750 --> 02:21.750 get to a point where more work happens, more analysis takes place so that you get a more comprehensive 02:21.750 --> 02:23.790 and more compelling outcome. 02:23.830 --> 02:28.990 Something which really can take several minutes and come back with something that is adding a lot of 02:28.990 --> 02:29.510 value. 02:29.630 --> 02:31.670 So that is the challenge for you. 02:31.790 --> 02:33.910 I would love to see your results. 02:34.030 --> 02:40.510 I'd love it if you could push your your projects to the community contributions folder and put your 02:40.510 --> 02:45.910 work into its own folder so that people can run your deep research agent separately. 02:45.950 --> 02:49.910 Perhaps you may need to do some fiddling around with gradio to be able to do things like surface those 02:49.910 --> 02:54.550 questions, but it's great fun to work with and it will be an awesome project. 02:54.550 --> 02:56.590 And it's also something that you can then showcase. 02:56.590 --> 02:58.510 You could because it's a Gradio app. 02:58.550 --> 03:04.680 You can of course type radio deploy to deploy it to hugging face, and that way you can have a spaces 03:04.680 --> 03:05.520 where it's live. 03:05.560 --> 03:11.920 You can share it on LinkedIn, and you can get others to try it out too, and be able to show people 03:11.920 --> 03:17.800 about the expertise that you've acquired in building sophisticated agentic AI. 03:19.480 --> 03:24.840 And with that, that wraps up week two, and I have had so much fun. 03:25.000 --> 03:31.760 You can tell I absolutely adore OpenAI agents SDK, and I'm excited to come back to it again in week 03:31.760 --> 03:32.240 six. 03:32.480 --> 03:36.120 But as it happens next week, we're working with Cru AI. 03:36.240 --> 03:43.040 And Cru AI is my second favorite, so it may not be quite quite as as dear in my heart as OpenAI agents 03:43.040 --> 03:44.640 SDK, but it's my second favorite. 03:44.680 --> 03:46.120 It's also really cool. 03:46.120 --> 03:51.080 It's very close to OpenAI agents SDK, and of course it has some different concepts, and there's some 03:51.080 --> 03:53.760 ways in which it's definitely better, as you will see. 03:53.840 --> 03:55.240 And I can't wait to show you. 03:55.240 --> 03:57.040 So see you for week three. 03:57.080 --> 03:59.240 Crew AI week coming right up.