WEBVTT 00:00.160 --> 00:00.720 Okay. 00:00.720 --> 00:05.240 And I will, uh, run the creating of the new servers and let's get on with it. 00:05.240 --> 00:09.240 So we're going to be creating two agents in this in this lab. 00:09.360 --> 00:12.320 One of them is called the researcher, and one of them is called the trader. 00:12.320 --> 00:16.840 And the researcher, of course, does market research on on the behest of the trader. 00:17.040 --> 00:19.320 And the trader makes trading decisions. 00:19.320 --> 00:24.320 And we're going to arm the researcher with tools to look online and both in the headless browser and 00:24.320 --> 00:26.520 by doing an actual search using brave. 00:26.960 --> 00:28.200 And now you might wonder why. 00:28.200 --> 00:34.000 Why not just simply give those two tools to the trader, just give all the tools to the trader, and 00:34.000 --> 00:38.800 let the trader do everything autonomously, and just have control of all of the different tools it uses. 00:38.840 --> 00:40.200 And you certainly can do that. 00:40.200 --> 00:46.320 And it might be worth experimenting with a benefit of separating out the researcher is it allows you 00:46.320 --> 00:52.880 to focus your prompting on one task to make sure that you don't distract the trader agent too much with 00:52.880 --> 00:59.960 too much optionality, almost so that there's some goal, some some tighter objective around research 00:59.960 --> 01:02.160 that we can set the researcher agent. 01:02.160 --> 01:04.440 And we also give it both of these tools. 01:04.480 --> 01:08.820 And if we were to give all of these tools to the trader agent, there's a risk that we confuse it by 01:08.820 --> 01:11.260 just giving it too many different tools that it could use. 01:11.380 --> 01:17.060 So this technique of trying to identify different steps, different stages, and then identifying an 01:17.060 --> 01:21.860 agent around it, it's a it's a great way to tackle a genetic projects. 01:22.140 --> 01:23.380 And it's experimental. 01:23.380 --> 01:25.140 There's no hard and fast rules. 01:25.140 --> 01:30.140 Sometimes it helps to give more flexibility to one agent with just a greater number of tools. 01:30.140 --> 01:35.900 Sometimes it helps to put boundaries around an activity and carve off an agent for that activity. 01:35.940 --> 01:37.420 And that's what we're doing right now. 01:38.220 --> 01:46.580 So the researcher agent we're saying the instructions for we're writing a coroutine get researcher instructions, 01:46.580 --> 01:48.020 which is the system prompt. 01:48.060 --> 01:49.620 You are a financial researcher. 01:49.620 --> 01:54.740 You're able to search the web for interesting news, look for possible trading opportunities and help 01:54.740 --> 01:56.940 with research based on a request. 01:56.940 --> 02:01.100 You carry out necessary research responding with your findings. 02:01.420 --> 02:08.500 Uh, and then, uh, we end by saying the current date time is and give the current date and time. 02:08.500 --> 02:15.130 It's just, again, a good a good tactic to arm the research agent with that information so that it 02:15.130 --> 02:16.290 doesn't think it's still. 02:16.330 --> 02:19.890 Its knowledge cut off date, which might be October of 2023 or something. 02:20.290 --> 02:24.650 So we're giving it the current date and that could be provided as a tool. 02:24.690 --> 02:29.850 But again, that would be a sort of needless distraction to confuse it by making it have to also call 02:29.850 --> 02:34.450 an extra tool, we might as well just shove that in the context every single time. 02:34.770 --> 02:41.530 And so with that, this will create an agent object an open AI agents SDK agent. 02:41.570 --> 02:42.970 The name is researcher. 02:43.010 --> 02:44.250 The instructions is. 02:44.250 --> 02:47.530 This system prompt will use GPT four mini for now. 02:47.530 --> 02:52.170 And we are passing in our MCP servers that get get brought in here. 02:52.170 --> 02:58.010 So we will pass them in to be the MCP servers for the researcher. 02:58.290 --> 02:59.770 And that I will run. 03:00.050 --> 03:08.090 And now I'm going to do this a get researcher tool, which is going to return the researcher converted 03:08.090 --> 03:14.850 into a tool using the as tool function that, uh, that OpenAI agents SDK gives us. 03:14.850 --> 03:15.890 Do you remember this? 03:15.890 --> 03:21.030 You remember that when you've got agents collaborating with the agents SDK, there's two ways of doing 03:21.030 --> 03:21.310 it. 03:21.350 --> 03:28.430 One is using handoffs, and the other is using as tool to convert an agent into a tool. 03:28.550 --> 03:32.230 And there are pros and cons and the different different use cases for using them. 03:32.510 --> 03:38.310 And the big distinction is that when it's a handoff the control is passed over to that other agent, 03:38.670 --> 03:44.070 and it's now delegated to the other agent to continue work and take it from there. 03:44.110 --> 03:46.950 When it's a tool, it's more of a request response. 03:47.310 --> 03:52.070 We will use that tool and the response will come back to the agent that initiated it. 03:52.110 --> 03:53.950 And that's the mode we're working on right now. 03:53.990 --> 04:00.310 We want the researcher to be something which is just only there for the purposes of helping the trader, 04:00.310 --> 04:03.950 and that's why it makes sense to convert the researcher to a tool. 04:04.230 --> 04:07.990 And you give it a tool name researcher and you describe the tool. 04:07.990 --> 04:11.430 This tool researches online for news and opportunities. 04:12.310 --> 04:15.310 Describe what kind of research you are looking for. 04:15.310 --> 04:17.950 So that would become the input to this tool. 04:18.590 --> 04:23.580 And if you need a refresher on that then just drop back to to week two when we had so much fun with 04:23.580 --> 04:25.820 both handoffs and tools. 04:25.860 --> 04:32.820 Okay, so in the past when we've used MCP servers, we've used something called a context manager, 04:32.820 --> 04:38.540 which is when you have stuff like, uh, you say um, uh, with. 04:41.140 --> 04:44.100 MCP studio Server, something like that. 04:44.100 --> 04:50.420 As server, when you use a context manager like that, what happens is that behind the scenes, the 04:50.460 --> 04:54.380 server, this server here is is called with connect. 04:54.420 --> 04:57.460 Server connect is called immediately when you do this. 04:57.500 --> 05:04.140 And then whenever it comes out of scope, uh, then automatically server cleanup is called which cleans 05:04.140 --> 05:07.140 up resources and stops connecting to the server. 05:07.420 --> 05:11.180 Uh, and uh, we are not going to do it that way. 05:11.180 --> 05:17.380 This time we're going to separately connect and later clean up, because we're going to want to distribute 05:17.380 --> 05:18.580 this over multiple cells. 05:18.580 --> 05:22.020 So we don't want to be constrained by having it in one context manager. 05:22.020 --> 05:23.340 But it's the same thing. 05:23.540 --> 05:31.160 So this here when we're going to say for all of the servers in our list of research servers await server 05:31.160 --> 05:32.320 dot connect. 05:32.440 --> 05:36.200 That's what happens automatically at the beginning of the with with statement. 05:37.200 --> 05:42.720 And then we are going to call our researcher passing in the research servers. 05:43.040 --> 05:51.240 And then we are going to call runner Dot saying what's the latest news on Amazon and then display the 05:51.240 --> 05:51.920 result. 05:51.920 --> 05:53.920 So this is us using this researcher. 05:53.920 --> 05:54.840 We're just going direct. 05:54.840 --> 05:56.080 We're not going through the tool. 05:56.080 --> 05:57.280 But it's the same idea. 05:57.280 --> 06:00.320 And we're asking for news on Amazon. 06:00.520 --> 06:02.160 Let's see how it does. 06:02.320 --> 06:03.440 It's thinking about that. 06:03.480 --> 06:05.240 It's no doubt showing some autonomy. 06:05.280 --> 06:11.200 Remember it has access to either run a headless browser or it can call the brave API. 06:11.640 --> 06:15.040 And it's got some news as of as of now. 06:15.040 --> 06:16.240 So it's got the current date. 06:16.240 --> 06:23.680 Well, the the bid to acquire TikTok, the spring sale success, new Prime Video content. 06:23.680 --> 06:29.000 So there's some some interesting stuff there that's coming from the news. 06:29.120 --> 06:31.880 So it shows that our research agent is working nicely. 06:31.880 --> 06:36.430 And again, it's the reason I like working in notebooks like this as we are iterating, because we're 06:36.430 --> 06:42.150 in this experimenters mindset where we try different things, explore different agents, make sure they're 06:42.150 --> 06:45.430 behaving the way we want before we move to a Python module. 06:45.430 --> 06:50.910 It's a great mode to work in as a scientist first and foremost before you become the engineer. 06:51.310 --> 06:56.550 But the other part of being a scientist is that we should come and look at the traits in, uh, in the 06:56.550 --> 06:57.310 platform. 06:57.510 --> 07:01.030 Openai.com going to traces to see what's happened. 07:01.030 --> 07:06.350 And there is a trace for researcher in we come and we can see what's going on here. 07:06.390 --> 07:09.390 It's, it's it's found out what MCP tools it has. 07:09.590 --> 07:11.510 It's then it's used the brave search. 07:11.550 --> 07:12.350 It's used it twice. 07:12.350 --> 07:13.950 It's done two different searches. 07:14.190 --> 07:18.430 Uh, latest news on Amazon and Amazon news April 2025. 07:18.430 --> 07:19.710 So a couple of different searches. 07:19.710 --> 07:20.310 It's done. 07:20.310 --> 07:21.710 It's got back the results. 07:21.710 --> 07:25.190 And it used that to craft its response that we already saw. 07:25.470 --> 07:27.470 Uh, so this is the trace. 07:27.470 --> 07:32.510 This is showing uh, what was what was going on that it was at work and, uh, yeah, that's that's 07:32.510 --> 07:33.830 all good to see. 07:34.230 --> 07:35.910 Um, okay. 07:36.670 --> 07:38.590 So back we go.