WEBVTT 00:00.150 --> 00:02.160 -: In the previous lesson, we looked at tool calling. 00:02.160 --> 00:03.570 So how to define a tool, 00:03.570 --> 00:06.270 how to use a tool within a large language model. 00:06.270 --> 00:07.470 In this video we're gonna have a look 00:07.470 --> 00:10.350 at how you can specifically create custom agents 00:10.350 --> 00:11.670 using these tools, 00:11.670 --> 00:13.260 putting them into a loop, 00:13.260 --> 00:15.600 and then breaking that loop accordingly. 00:15.600 --> 00:17.970 The first thing I want you to do is open the notebook 00:17.970 --> 00:20.220 in the openai_features_and_functionality folder 00:20.220 --> 00:22.801 called learning_agents_from_scratch. 00:22.801 --> 00:25.097 So what we're gonna do is import openai, 00:25.097 --> 00:27.420 numpy, matplotlib, and pydantic. 00:27.420 --> 00:31.050 We're gonna have a look at is using a agentic loop 00:31.050 --> 00:32.430 that we'll build from scratch, 00:32.430 --> 00:34.260 and you can either copy 00:34.260 --> 00:36.600 or you can follow along and code it from scratch. 00:36.600 --> 00:38.340 The thing we're gonna be coded from scratch 00:38.340 --> 00:41.160 is a simple agent that's in a while true loop. 00:41.160 --> 00:44.730 So firstly you need to import your OpenAI package, 00:44.730 --> 00:48.780 and the model that we'll be using is gpt-4.1-mini. 00:48.780 --> 00:51.930 Now, we've already set up the get_weather function 00:51.930 --> 00:53.610 and the weather tool for you. 00:53.610 --> 00:56.220 You're going to need to replace your OpenAI key here, 00:56.220 --> 00:57.510 so I'll just let you do that. 00:57.510 --> 00:59.460 Okay, so we have the get_weather function, 00:59.460 --> 01:01.980 we have a weather_tool dictionary here, 01:01.980 --> 01:04.860 and we also have our tools Python list that we've set up 01:04.860 --> 01:07.530 with the information for a single weather tool. 01:07.530 --> 01:09.990 Now, if you scroll down, here's where it gets interesting. 01:09.990 --> 01:11.820 So we're gonna set our messages up 01:11.820 --> 01:13.950 and what we're gonna do for our messages 01:13.950 --> 01:15.990 is we're gonna start with one message. 01:15.990 --> 01:18.633 So I'm gonna put role as equal to user, 01:19.740 --> 01:21.990 and the content is, "What's the weather like 01:21.990 --> 01:23.430 in Paris today?" 01:23.430 --> 01:24.930 Let's get rid of this section here. 01:24.930 --> 01:27.510 So I'm just gonna do just the weather in Paris. 01:27.510 --> 01:29.040 And then what we need to do 01:29.040 --> 01:31.590 is firstly we create our initial response, 01:31.590 --> 01:33.360 given this message, 01:33.360 --> 01:36.840 and we have the model is equal to our gpt-4.1 model. 01:36.840 --> 01:38.700 The input is equal to the messages 01:38.700 --> 01:41.700 and our tools is equal to the tools that we defined earlier. 01:41.700 --> 01:44.130 Then what we're gonna do is process all function calls 01:44.130 --> 01:45.300 in the response to output. 01:45.300 --> 01:48.150 So if there's a response to output, 01:48.150 --> 01:50.880 what we're then gonna do is we're gonna loop through 01:50.880 --> 01:52.620 all of the output items. 01:52.620 --> 01:54.150 So you'll see here, we're going 01:54.150 --> 01:57.180 for output_item in response to output. 01:57.180 --> 01:59.580 If we have a .type 01:59.580 --> 02:01.530 and it's equal to a function call, 02:01.530 --> 02:04.080 what we're then gonna do is append the function call 02:04.080 --> 02:05.400 to the messages. 02:05.400 --> 02:07.470 After that, we're gonna get the tool call 02:07.470 --> 02:10.642 so that would be tool call = output_item 02:10.642 --> 02:12.810 and then we're gonna load in the args, 02:12.810 --> 02:14.410 json.loads(tool_call.arguments). 02:16.620 --> 02:18.990 After that, we're going to execute the function. 02:18.990 --> 02:20.187 So you've got this get_weather 02:20.187 --> 02:23.340 and we're passing in the args of latitude and longitude 02:23.340 --> 02:24.720 and we're printing out the fact 02:24.720 --> 02:26.610 that we've executed that tool call. 02:26.610 --> 02:27.720 And then what we're gonna do 02:27.720 --> 02:30.270 is now that we've executed that tool call, 02:30.270 --> 02:34.530 we need to then append the messages back into the LLM. 02:34.530 --> 02:36.783 So that'll be type function_call_output. 02:36.783 --> 02:40.380 The call_id is equal to the tool_call.call_id, 02:40.380 --> 02:42.900 and the output is the string of this result. 02:42.900 --> 02:45.480 Now, this is similar to what you have with function calling. 02:45.480 --> 02:47.070 The only difference is, 02:47.070 --> 02:48.450 what we're now gonna check for 02:48.450 --> 02:51.120 is if there's any output text. 02:51.120 --> 02:54.150 So we're gonna say if there's any output text, 02:54.150 --> 02:56.283 and we'll do this out here, 02:58.380 --> 03:00.120 if there's any output text, 03:00.120 --> 03:03.300 then we have a final output and we have a break. 03:03.300 --> 03:05.207 And if there isn't a response to output, 03:05.207 --> 03:06.720 we are then just gonna break 03:06.720 --> 03:09.300 just to be on the safe side of things. 03:09.300 --> 03:10.770 So what's gonna happen is, 03:10.770 --> 03:12.810 we have this client.responses.create, 03:12.810 --> 03:15.150 we have a query that's gonna trigger a tool call, 03:15.150 --> 03:16.230 if there's an output, 03:16.230 --> 03:17.637 we're gonna loop through every output 03:17.637 --> 03:19.080 and if it's a function call, 03:19.080 --> 03:21.450 then we're gonna add that tool call to the message. 03:21.450 --> 03:22.830 We're then gonna get that tool call, 03:22.830 --> 03:24.750 get the various bits of information, 03:24.750 --> 03:26.820 we're gonna get the weather and invoke that. 03:26.820 --> 03:29.190 We're then gonna tell the messages 03:29.190 --> 03:31.410 that this is the output of that tool call. 03:31.410 --> 03:33.450 And then this is the clever stuff here 03:33.450 --> 03:36.360 where you can basically say, if we have an output text, 03:36.360 --> 03:38.670 then that basically means we're finished at this point 03:38.670 --> 03:40.890 and we'll break out of the endless while loop. 03:40.890 --> 03:42.270 Now, if we didn't have an output before, 03:42.270 --> 03:43.680 just for simplicity, 03:43.680 --> 03:45.960 then we would also add a break here. 03:45.960 --> 03:46.980 So if you run this now, 03:46.980 --> 03:49.770 what you'll see is if you scroll up here, 03:49.770 --> 03:51.840 it's gonna execute that tool call 03:51.840 --> 03:53.550 and it returns the final output. 03:53.550 --> 03:55.440 Now, we can take this a step further 03:55.440 --> 03:58.050 and just wrap all the code that you just wrote 03:58.050 --> 04:00.360 into a python function called an agent_loop, 04:00.360 --> 04:02.820 which just takes in messages and tools. 04:02.820 --> 04:05.100 And then we can try and run that on a query 04:05.100 --> 04:07.890 that has two different types of intents. 04:07.890 --> 04:10.260 So you'll see here we've got a role of developer, 04:10.260 --> 04:12.990 that's basically the same as the system message, 04:12.990 --> 04:16.410 which is useful for instructional pieces of content. 04:16.410 --> 04:18.210 Then we have a content key, 04:18.210 --> 04:19.417 and under that it says, 04:19.417 --> 04:21.480 "What's the weather like in Paris today? 04:21.480 --> 04:22.980 Before applying, I want you to also 04:22.980 --> 04:24.510 get the weather for Berlin." 04:24.510 --> 04:26.580 Now, when we run our agent_loop function, 04:26.580 --> 04:28.050 what you'll see is it actually decides 04:28.050 --> 04:31.230 to execute two tool calls before responding with the output. 04:31.230 --> 04:34.230 So again, the important point is we put the agent 04:34.230 --> 04:36.330 into a while true loop 04:36.330 --> 04:37.770 and we loop through 04:37.770 --> 04:40.140 and anytime we have tool calls, we're continuing, 04:40.140 --> 04:42.540 but if we have that .output_text property, 04:42.540 --> 04:44.040 then we're breaking out of the loop. 04:44.040 --> 04:45.780 You can also do it slightly differently 04:45.780 --> 04:47.730 with a objective function. 04:47.730 --> 04:49.200 And basically what that means is you have 04:49.200 --> 04:50.700 some kind of custom logic 04:50.700 --> 04:52.470 where rather than just endlessly being 04:52.470 --> 04:54.780 in a while true loop, you might decide, 04:54.780 --> 04:59.250 okay, the agent has a maximum of five times that it can run 04:59.250 --> 05:01.500 or a certain amount of time that it can run. 05:01.500 --> 05:03.780 In our case, we have an objective function 05:03.780 --> 05:06.840 with this objective met in terms of the search count, 05:06.840 --> 05:09.510 is it greater than the maximum number of searches? 05:09.510 --> 05:10.770 We can then do something like this 05:10.770 --> 05:12.960 where the role of developer, you know, 05:12.960 --> 05:14.017 you have the content, 05:14.017 --> 05:15.450 "Your goal is to gather weather 05:15.450 --> 05:17.160 for at least five different cities. 05:17.160 --> 05:19.710 Once you've done that, respond with task complete." 05:19.710 --> 05:21.780 And then what you can see is we have this search, 05:21.780 --> 05:23.100 the weather in Berlin. 05:23.100 --> 05:25.680 Now, we can still use a while true, 05:25.680 --> 05:29.130 but what we can do then is we can then basically do 05:29.130 --> 05:30.630 the same kind of logic, 05:30.630 --> 05:33.240 however we increment the search count. 05:33.240 --> 05:35.400 So then what we do is check if the objective is met 05:35.400 --> 05:36.630 based on that search count, 05:36.630 --> 05:39.990 and if necessary, then we will break out of the loop. 05:39.990 --> 05:42.030 So you can see here, when we run this, 05:42.030 --> 05:44.280 what will happen is we'll be getting weather. 05:46.980 --> 05:48.570 So it's decided to do New York 05:48.570 --> 05:49.680 and it's getting some weather 05:49.680 --> 05:52.260 and it's then it's searched in six cities 05:52.260 --> 05:55.320 and after it's searched over a certain number, 05:55.320 --> 05:57.360 it has exited the while true loop 05:57.360 --> 05:59.940 because of this objective_met function. 05:59.940 --> 06:02.160 So basically all you're doing here is creating 06:02.160 --> 06:05.100 a custom function so that it doesn't endlessly go around 06:05.100 --> 06:06.150 in a while true loop. 06:06.150 --> 06:10.110 It can exit before, by using a custom objective function. 06:10.110 --> 06:12.030 The other common ways to do this 06:12.030 --> 06:14.740 is have, like, a MAXIMUM_NUMBER_OF_ITERATIONS 06:16.321 --> 06:18.450 and then you can then say, 06:18.450 --> 06:20.790 when you're doing these kind of iterations, 06:20.790 --> 06:22.560 you will then have something like this 06:22.560 --> 06:25.383 where if you are over the number of iterations, 06:26.354 --> 06:28.533 we also have to store the current iterations, 06:28.533 --> 06:30.960 CURRENT_NUMBER_oF_ITERATIONS, 06:30.960 --> 06:32.880 if the current number of iterations 06:32.880 --> 06:33.960 is greater than or equal to 06:33.960 --> 06:36.150 the maximum number of iterations, 06:36.150 --> 06:39.300 then we're breaking out of the loop. 06:39.300 --> 06:40.290 And you can see here, 06:40.290 --> 06:41.280 every time we go around, 06:41.280 --> 06:43.950 we're incrementing the current number of iterations. 06:43.950 --> 06:46.230 So this is another way that you could also break 06:46.230 --> 06:47.280 out that while loop 06:47.280 --> 06:49.260 so that your agent doesn't endlessly go on 06:49.260 --> 06:52.620 if it doesn't meet a custom objective function. 06:52.620 --> 06:54.210 Okay, just to recap, 06:54.210 --> 06:56.220 we looked at how we could use the while true loop 06:56.220 --> 06:59.280 to basically create an agentive while loop, 06:59.280 --> 07:00.570 and then we looked at the fact 07:00.570 --> 07:02.220 that we're doing function calling here 07:02.220 --> 07:04.590 and we also have a custom step 07:04.590 --> 07:07.230 which will break us out of the whiled true loop. 07:07.230 --> 07:10.290 This is how a lot of the agents under the hood are designed 07:10.290 --> 07:12.360 and it's well worthwhile, you getting familiar 07:12.360 --> 07:13.920 with this type of pattern. 07:13.920 --> 07:15.390 Secondly, we looked at the fact 07:15.390 --> 07:18.480 that agents can do parallel tool calling. 07:18.480 --> 07:19.770 This mixed intent query 07:19.770 --> 07:21.720 is searching for two types of weather 07:21.720 --> 07:24.420 and we can see that it executed two tool calls 07:24.420 --> 07:26.820 before responding to the user. 07:26.820 --> 07:30.630 Thirdly, we looked at making custom objective functions 07:30.630 --> 07:33.300 and using an objective function 07:33.300 --> 07:35.490 or the maximum number of iterations 07:35.490 --> 07:37.530 to break out of a while loop 07:37.530 --> 07:40.170 to avoid the fact that these agents could get stuck 07:40.170 --> 07:41.520 for larger periods of time. 07:41.520 --> 07:43.270 Cool, I'll see you in the next one.