WEBVTT 00:00.800 --> 00:04.280 And let's now go and rerun everything. 00:12.200 --> 00:17.520 And here we see that we initialize the session, we process the request of list tools. 00:17.520 --> 00:20.720 And then we had a request of call tool. 00:20.880 --> 00:29.240 So our graph agent our graph react agent which is our MCP host use the MCP client in order to invoke 00:29.640 --> 00:33.000 the add tool that the MCP server exposes. 00:34.320 --> 00:42.520 And if I will go to the MCP flow of the protocol, then the query we send to our app host. 00:42.680 --> 00:49.360 Then it sends to the LM the same query plus all the information about the MCP tools. 00:49.680 --> 00:57.520 So then the LM responded that it should go and run the add function with the correct arguments. 00:57.520 --> 00:59.240 So it's going to be two two. 01:00.200 --> 01:02.280 And something very cool happened here. 01:02.560 --> 01:10.240 So instead of the agent running and executing the tools in the actual graph application, what's happening 01:10.240 --> 01:18.280 is that the graph agent, with the help of the MCP client, is making a request to the MCP server, 01:18.480 --> 01:24.840 and the actual execution of the tool is happening in the server side, which is decoupled from our graph 01:24.840 --> 01:25.800 application. 01:25.800 --> 01:29.240 So our graph application is responsible for the orchestration. 01:29.440 --> 01:33.640 And the MCP server is responsible for the execution of the tools. 01:33.680 --> 01:38.560 This is one of the benefits of combining land graph and link chain with MCP. 01:40.080 --> 01:40.760 Alrighty. 01:40.800 --> 01:49.080 So after we execute it in the server side, the tool we send back the response to the MCP client which 01:49.160 --> 01:57.040 forwards it into our graph agent, which then sends the original query with the tool execution answer 01:57.040 --> 01:58.000 to the LM. 01:58.280 --> 02:03.400 The LM is then going to process it, and it's going to output as the final answer in this case. 02:03.600 --> 02:08.000 And this is what's being forwarded to the user by our application. 02:08.440 --> 02:10.840 So let's go back here to the final answer. 02:10.840 --> 02:13.880 And the final answer is two plus two is equal to four. 02:15.560 --> 02:18.120 And let me now go and change here instead of two. 02:18.160 --> 02:20.560 Let's put here for example 54. 02:20.920 --> 02:23.560 And let's make this example a bit more complicated. 02:23.560 --> 02:26.040 And let's add here times three. 02:26.840 --> 02:30.160 So we have here two invocations of the tool here. 02:30.640 --> 02:33.200 So let's go and rerun everything. 02:36.680 --> 02:39.360 And we can see we actually get three tool calls. 02:39.360 --> 02:43.280 And for me it's a surprise because I thought we would get two one call. 02:43.280 --> 02:46.880 It's going to be two times three which is going to output six. 02:47.000 --> 02:50.320 And the other two will call is going to be six plus 54. 02:50.600 --> 02:54.360 So we'll check in the trace soon and we'll see soon what's happening here. 02:54.560 --> 02:56.080 And let's wait for the result. 02:56.120 --> 02:57.720 The result here is 60. 02:57.840 --> 02:59.800 So the answer here is correct. 02:59.800 --> 03:03.640 So to complete the picture I want to go into lengthy tracing. 03:03.640 --> 03:10.030 And I want to see exactly what were the tool calls and to see what was this mysterious extra tool call. 03:10.470 --> 03:11.230 Alrighty. 03:11.390 --> 03:16.590 So let me open links meat and here I will go to MCP test. 03:16.630 --> 03:18.670 This is the project I'm tracing right now. 03:19.790 --> 03:24.190 And here we have two traces of our graph executions. 03:24.350 --> 03:27.390 So one is going to be for the two plus two query. 03:27.390 --> 03:31.070 And the other is going to be for the 54 plus two times three query. 03:31.550 --> 03:37.870 By the way I made those traces public and I am sharing them in the videos resources. 03:37.870 --> 03:40.390 So you can see exactly what I am seeing right now. 03:40.510 --> 03:47.870 If you didn't do it for yourself, let's open the two plus two trace and let's examine what are the 03:47.870 --> 03:49.350 calls made to the LM. 03:49.870 --> 03:54.630 And let's go and select here this instance of chat OpenAI message. 03:55.070 --> 03:56.750 And here we're seeing the trace. 03:57.230 --> 04:01.070 And we can see that we made a request with function calling. 04:01.230 --> 04:03.230 So we can see here the tools we have here. 04:03.230 --> 04:07.030 This is the tools that the graph agent sent. 04:07.230 --> 04:10.070 And we can see we have here with the add and then multiply tool. 04:10.230 --> 04:15.110 And we can see that the LM responded with the add tool that needs to be called. 04:15.110 --> 04:17.710 And we can see it by this called tag over here. 04:17.710 --> 04:20.030 So the input was what is two plus two. 04:20.070 --> 04:22.750 So this was the exact input of the original query. 04:23.070 --> 04:27.750 The augmentation with the tools that Landgraaf did for us was here. 04:28.030 --> 04:32.390 And the result is this tool call with those arguments here. 04:33.030 --> 04:39.590 So the MCP client then made a request to the MCP server with the input of two plus two. 04:39.630 --> 04:42.230 So we executed the tool in the MCP server. 04:42.430 --> 04:45.030 And this is the tracing of this execution. 04:45.030 --> 04:49.190 And we can see here the input is two and two and the output is four. 04:49.550 --> 04:56.670 And something interesting to note is that in this trace here we can't really see that the tool was actually 04:56.670 --> 05:02.710 executed on the MCP server, which is completely decoupled from the graph graph. 05:03.110 --> 05:06.870 So we can't really see it even by inspecting the metadata here. 05:07.070 --> 05:12.710 But I'm positive that the LinkedIn team is going to update this, and it's going to make this more visible 05:12.710 --> 05:17.430 and more noticeable for agents which leverage MCP servers. 05:19.110 --> 05:22.150 Anyways, here we can see now the final answer. 05:22.150 --> 05:26.190 So we have here the original prompt the tool call. 05:26.190 --> 05:28.750 And we get here the final answer which is four. 05:30.070 --> 05:36.350 And the last thing I owe you for this session is to show you this extra mysterious tool call that we 05:36.350 --> 05:37.790 had in the second round. 05:38.910 --> 05:42.590 And if we'll open the first request to the LLM. 05:42.590 --> 05:47.710 So this is the original prompt alongside the augmentation of the MCP tools here. 05:47.990 --> 05:53.270 Then we can see that the output we get back from the LLM is a multiple tool call. 05:53.270 --> 05:54.990 So we have here two tool calls. 05:54.990 --> 05:57.590 One is going to be 54 plus zero. 05:57.750 --> 06:01.430 And this is the extra tool call that I missed. 06:01.590 --> 06:04.310 And the other is going to be two times three. 06:04.590 --> 06:10.340 And an interesting thing with graph by the way is that it's going to run those tool calls concurrently. 06:10.340 --> 06:13.940 So it's going to make two concurrent requests into our MCP server. 06:13.940 --> 06:17.260 So one request is going to be to add 54 and zero. 06:17.500 --> 06:20.900 And the other request is going to be to multiply two and three. 06:21.100 --> 06:26.460 And you can see that they are concurrent because they are starting in the same start time, the same 06:26.460 --> 06:30.020 timestamp of 1131 and 17 milliseconds. 06:32.060 --> 06:36.660 And the rest of the execution is pretty straightforward. 06:36.900 --> 06:43.340 After we have the result from the MCP server of how much is two times three, we can send another request 06:43.380 --> 06:46.460 to the Lem with this observation with this result. 06:46.620 --> 06:48.980 So let's go and check it out. 06:50.300 --> 06:57.180 And here we are sending now the Lem, the same query but with the result of the tool call. 06:57.580 --> 07:01.940 And then the Lem is going to decide that it needs to add up 54 and six. 07:02.140 --> 07:04.620 So let's go and see if this tool execution here. 07:08.060 --> 07:10.540 And the rest is pretty straightforward. 07:10.540 --> 07:15.980 We send this final result to the LM again, alongside with the history of what we did. 07:16.020 --> 07:22.100 And then we get the final answer and let's go and let's check out the final answer here. 07:23.620 --> 07:25.220 And that's it. 07:25.260 --> 07:32.620 Let me go now and go back to cursor and let me go and add the implementation here. 07:32.620 --> 07:33.980 And let's commit it to GitHub. 07:34.660 --> 07:40.420 So I'm going to use again the cursor feature of creating the commit message. 07:44.340 --> 07:49.340 Let me commit that and let me push to the repository. 07:56.140 --> 07:59.420 All right so you can see the final artifact. 07:59.620 --> 08:02.580 Let's go to the GitHub repository. 08:02.860 --> 08:06.220 So the final artifact is going to be here in this branch. 08:06.420 --> 08:09.820 And you can see here that this is the code.