WEBVTT 00:00.040 --> 00:02.960 Well, I don't know how this day always comes along so fast. 00:02.960 --> 00:07.800 It feels like just the other day we were starting, we were doing instant gratification and then the 00:07.800 --> 00:08.920 foundations. 00:08.920 --> 00:10.120 And now here we are already. 00:10.120 --> 00:11.840 It's the penultimate day. 00:12.040 --> 00:14.280 It is week three, day four. 00:14.320 --> 00:15.800 It's our last Purple day. 00:16.080 --> 00:19.720 As we embark on the topic of context engineering. 00:19.960 --> 00:22.840 Subagents so much juicy stuff. 00:22.840 --> 00:29.560 And then we begin today, the capstone project that we will complete tomorrow as we wrap the whole thing 00:29.560 --> 00:29.960 up. 00:30.480 --> 00:31.280 Let's get going. 00:31.280 --> 00:35.320 I promised you a big day yesterday and it was a fiendishly big day. 00:35.440 --> 00:44.640 We covered MCP, the model context protocol from anthropic, as they call it, the USB-C for AI apps, 00:44.680 --> 00:51.240 and they also call it the open source standard that connects AI applications to external systems. 00:51.240 --> 00:55.360 Or really, for the part that we really care about is tools. 00:55.560 --> 00:59.200 And I started yesterday by by going through what MCP is not. 00:59.280 --> 01:04.780 And after we laid that to rest, we covered the couple of things that MCP really is, and we put both 01:04.780 --> 01:05.780 of them to use. 01:05.980 --> 01:10.980 And of course, one of them and the main one, the one that is most important and the one that you should 01:10.980 --> 01:18.740 definitely remember, is that MCP gives us this simple way to use tools that have been written by somebody 01:18.740 --> 01:19.380 else. 01:19.540 --> 01:24.660 And so if someone has written MCP tools and they've packaged them into this, this MCP way of doing 01:24.660 --> 01:31.140 things, then it's really easy to connect them in to your app or your workflow. 01:31.300 --> 01:37.420 So it turns out that there's some things where Nan did not have a pre-baked tool that we could just 01:37.420 --> 01:44.660 use, Fire Crawl in particular, but Fire Crawl did have an MCP server, and so we could create an MCP 01:44.660 --> 01:50.820 client node that connected directly to it and just exposed its tools to our LLM. 01:51.060 --> 01:52.460 And it worked great. 01:52.620 --> 02:00.100 And we did it for Hunter.io as well, even though Hunter.io does have a tool in in the library, but 02:00.100 --> 02:03.580 it turns out that it was able to use some tools that aren't exposed to Nan. 02:03.620 --> 02:06.530 So we did get some benefit anyway by using MCP. 02:06.730 --> 02:08.170 So it was it was pretty cool. 02:08.210 --> 02:13.370 But the takeaway here is that this is the way that people use MCP, and this is the thing you need to 02:13.410 --> 02:14.370 know about that. 02:14.410 --> 02:19.490 MCP gives you an ability to use someone else's tools in a really easy way. 02:19.610 --> 02:21.090 As we saw for ourselves. 02:21.250 --> 02:26.890 And then the other thing that you can do with MCP is just the flip of this is to say, if you've built 02:26.890 --> 02:31.410 a tool and you like the tool you've built, and it's something that other people could benefit from, 02:31.450 --> 02:37.850 or maybe you could benefit from it in a different app, then MCP gives you this really easy way to equip 02:37.890 --> 02:41.770 another LLM with your tool to share your tool. 02:41.890 --> 02:44.010 And that is also what we did. 02:44.010 --> 02:45.010 We took everything. 02:45.010 --> 02:49.810 We'd built our prospecting subagent, which itself was using MCP. 02:49.890 --> 02:51.970 It didn't need to of course, it could have been anything. 02:51.970 --> 02:58.170 But we happened to pick that that prospecting agent and we packaged all of that up into an MCP server. 02:58.210 --> 03:04.850 We added an MCP server trigger node, and then we just said, okay, this is MCP server. 03:04.850 --> 03:12.150 And that means that anyone that has that URL can use This can can equip their Lem to use this tool as 03:12.150 --> 03:14.630 part of their workflow and to try it out. 03:14.670 --> 03:18.470 We picked something I picked clawed the the application. 03:18.470 --> 03:20.750 It's anthropic's version of ChatGPT. 03:20.990 --> 03:28.590 We use that, we connected it to our prospecting MCP server, and suddenly Claude was able to hunt for 03:28.590 --> 03:33.030 leads for us using behind the scenes using Fire Crawl and Hunter.io. 03:33.230 --> 03:38.430 Not that it knew that, but all of that was happening and I thought that was just a great end to end 03:38.470 --> 03:39.190 example. 03:39.190 --> 03:47.150 We were sharing tools through MCP, and as it happens, those tools themselves were using MCP servers 03:47.150 --> 03:47.670 too. 03:47.710 --> 03:49.990 They were using Fire Crawl and Hunter.io. 03:49.990 --> 03:56.390 So we were using MCP at two levels, and I thought that was just a just a great end to end example, 03:56.470 --> 03:58.150 and I hope it landed for you too. 03:58.190 --> 04:05.110 And then one more thing to recap on is structured outputs that honestly, I think this is so important. 04:05.150 --> 04:09.030 And they don't get enough air time and they should do structured outputs. 04:09.030 --> 04:15.300 Of course it is this idea that you can say an LLM that normally just generates natural language, you 04:15.300 --> 04:16.500 can constrain it. 04:16.500 --> 04:24.220 So it has to generate JSON and that JSON needs to to fit to a particular structure called the schema. 04:24.460 --> 04:31.300 And you do this in N by turning on that switch that says require specific output format. 04:31.460 --> 04:35.540 And then you can give it a JSON example as I've given here. 04:35.740 --> 04:41.140 And when the LLM generates output it will need to conform to that structure. 04:41.140 --> 04:48.260 And it gives you this incredibly powerful way to put some kind of controls around what is otherwise 04:48.260 --> 04:50.140 very free flowing text. 04:50.140 --> 04:52.180 And you can use that in all sorts of ways. 04:52.180 --> 04:57.820 And in N-r10, it's particularly useful because you can use it to decide what to do next, knowing that 04:57.820 --> 05:00.380 you're going to get an output with a particular structure. 05:00.420 --> 05:03.140 Now, here's an interesting and important point. 05:03.300 --> 05:09.540 You will often find as you build workflows in N-r10, that there are actually two different ways that 05:09.540 --> 05:11.020 you could achieve the same thing. 05:11.180 --> 05:18.560 You could use tools, or alternatively, you could have your AI agent generate structured outputs and 05:18.560 --> 05:22.160 then interpret that output and make some decision as a result. 05:22.160 --> 05:26.880 And I think we did something a bit like that the first time we worked with Fire Crawl, we had that 05:26.880 --> 05:28.840 being and we outputted JSON. 05:28.840 --> 05:32.120 And then that went to some decisions that that called Fire Crawl. 05:32.120 --> 05:34.520 And then later we used a tool. 05:34.840 --> 05:37.320 And you'll often have those two different approaches. 05:37.320 --> 05:40.000 And you may end up thinking, I wonder which one is better. 05:40.280 --> 05:44.520 And you know, the answer is, of course, they have pros and cons, as with with many things. 05:44.520 --> 05:50.400 But but I will tell you, in my opinion, structured outputs is often the one to choose. 05:50.600 --> 05:56.560 It's the more boring answer tools is, the more the more trendy approach. 05:56.560 --> 05:59.520 It's the one where you're giving the most autonomy to your model. 05:59.560 --> 06:00.320 It's optional. 06:00.320 --> 06:04.160 Whether it uses the tool or not, it can have multiple tools. 06:04.160 --> 06:05.720 It can use any of them. 06:05.840 --> 06:08.120 It can choose what order it uses a tool in. 06:08.120 --> 06:09.960 It can call a tool multiple times. 06:10.160 --> 06:12.960 So there's much more flexibility with tools. 06:13.280 --> 06:18.630 But if you describe everything that you want Clearly in a JSON structure. 06:18.630 --> 06:23.790 And then you take that JSON structure and you use that to drive whatever happens next. 06:23.830 --> 06:30.270 That is a more kind of regimented, a more fixed way of doing things where you get to control how flexible 06:30.270 --> 06:33.590 it is or how inflexible it is in a lot of detail. 06:33.590 --> 06:40.710 And so I would describe the structured outputs alternative as the boring but bulletproof approach. 06:40.830 --> 06:46.950 And what I've built these things in production I have to I have to come clean that I almost always use 06:46.950 --> 06:53.190 the structured output style rather than too many tools, because you need something that is reliable. 06:53.310 --> 06:58.590 I would prioritize bulletproof over being being fashionable, so it's something to leave you with. 06:58.630 --> 07:03.230 You'll often find you have to make that decision, and you're wondering and you'll see what I mean about 07:03.270 --> 07:04.390 it being boring. 07:04.430 --> 07:06.470 You often have to do a bit more build work. 07:06.510 --> 07:07.830 It's quite repetitive. 07:07.950 --> 07:16.430 It's it reduces some of the magic of letting an AI agent choose its tools, but it gives you reliability 07:16.430 --> 07:17.510 as a result of that. 07:17.510 --> 07:19.950 And for me, that's the top priority.