1 00:00:00,200 --> 00:00:00,760 Okay. 2 00:00:01,280 --> 00:00:02,080 Moment of truth. 3 00:00:02,080 --> 00:00:02,720 I'm going to run it. 4 00:00:02,720 --> 00:00:04,880 And while it's running, I will talk through what it's doing. 5 00:00:04,920 --> 00:00:05,320 Hang on. 6 00:00:05,360 --> 00:00:06,080 Off it goes. 7 00:00:06,400 --> 00:00:10,640 So we have a new sales manager. 8 00:00:10,640 --> 00:00:14,600 We say you are the sales manager working for comply. 9 00:00:14,760 --> 00:00:17,600 You use the tools given to you to generate cold sales emails. 10 00:00:17,640 --> 00:00:18,760 Never generate one yourself. 11 00:00:18,800 --> 00:00:19,840 Always use the tools. 12 00:00:19,840 --> 00:00:24,440 You try all three sales email tools at least once before choosing the best one. 13 00:00:24,440 --> 00:00:26,560 You can use all the tools multiple times. 14 00:00:26,560 --> 00:00:31,800 If you're not satisfied with the results from the first try, you select the single best email using 15 00:00:31,800 --> 00:00:38,120 your own judgment of which email will be most effective, and then after picking it, you hand off to 16 00:00:38,160 --> 00:00:42,200 the email manager agent to format and send the email. 17 00:00:42,400 --> 00:00:47,320 Now again, I'm really spelling it out here, and you shouldn't need to be that prescriptive because 18 00:00:47,320 --> 00:00:49,640 the hand off description was there. 19 00:00:49,640 --> 00:00:51,880 But it helps to be to be precise. 20 00:00:51,920 --> 00:00:56,600 And particularly as these things get more complex, if you want it to stay on track and to be reliable, 21 00:00:56,600 --> 00:00:59,080 then this is a good practice to do. 22 00:00:59,440 --> 00:01:01,720 So then we create the agent. 23 00:01:01,960 --> 00:01:03,120 We it's a sales manager. 24 00:01:03,120 --> 00:01:04,480 We give it these instructions. 25 00:01:04,480 --> 00:01:08,470 We pass in both the tools and the handoffs. 26 00:01:08,470 --> 00:01:10,310 This is the crux of this. 27 00:01:10,510 --> 00:01:15,430 This is how we separate out that the tools, which is where we it's like a request response from the 28 00:01:15,470 --> 00:01:21,830 handoffs, which is a delegation control is passed across to take care of the rest of the work. 29 00:01:22,230 --> 00:01:24,910 And that is our message. 30 00:01:25,190 --> 00:01:30,630 And we put a trace around it, automated SDR and we give it a run. 31 00:01:30,630 --> 00:01:34,510 And it took about a minute and let's see what came back. 32 00:01:34,990 --> 00:01:39,790 This is the result and it's picked a very short one in the end. 33 00:01:39,790 --> 00:01:42,750 You remember one of them is a very concise email writer. 34 00:01:42,750 --> 00:01:44,390 It does indeed begin, dear CEO. 35 00:01:44,390 --> 00:01:46,270 And it ends, Alice, as we asked for. 36 00:01:46,510 --> 00:01:51,830 And it's given a nice, short, sharp, AI driven solutions that can optimize your business processes. 37 00:01:52,150 --> 00:01:55,230 And so it's nice and crisp. 38 00:01:55,710 --> 00:01:56,710 There we go. 39 00:01:57,510 --> 00:02:01,430 And importantly, we now have to go and look at the trace. 40 00:02:01,630 --> 00:02:04,790 So the trace here it is automated SDR. 41 00:02:04,830 --> 00:02:06,270 You can see how long it took. 42 00:02:06,390 --> 00:02:08,350 It used nine tools. 43 00:02:08,510 --> 00:02:09,750 Let's go and have a look. 44 00:02:09,870 --> 00:02:14,460 So it came in and it used sales agent one and two and three. 45 00:02:14,620 --> 00:02:15,740 And then look at what happened. 46 00:02:15,740 --> 00:02:19,660 It went back and used sales one, two and three a second time. 47 00:02:19,820 --> 00:02:20,540 Load more. 48 00:02:21,740 --> 00:02:26,100 And then this was the handoff you can see here shown. 49 00:02:26,340 --> 00:02:28,500 Then it goes to the email manager. 50 00:02:28,700 --> 00:02:34,460 And then that goes to the subject writer followed by the converter followed by the send email. 51 00:02:34,460 --> 00:02:37,460 And you'll see the control does not then come back. 52 00:02:37,780 --> 00:02:40,580 So you're seeing everything in action here. 53 00:02:40,580 --> 00:02:43,740 The sales manager that's able to to handle things. 54 00:02:43,740 --> 00:02:44,980 And it hands off here. 55 00:02:44,980 --> 00:02:51,180 And the email manager takes the rest of control for the rest of this blue timeline that you see highlighted 56 00:02:51,180 --> 00:02:51,780 in here. 57 00:02:51,780 --> 00:02:58,940 You can see the multiple calls to different agents and the handoff and the use of tools, both wrapping 58 00:02:58,940 --> 00:03:01,940 agents and just directly calling functions. 59 00:03:02,220 --> 00:03:10,540 And that is a satisfying conclusion to quite an interesting example of a genetic workflows and design 60 00:03:10,540 --> 00:03:11,180 patterns. 61 00:03:11,860 --> 00:03:12,380 Okay. 62 00:03:12,420 --> 00:03:19,810 Well, we covered a few of the constructs and concepts with OpenAI agents SDK, including tools and 63 00:03:19,810 --> 00:03:20,490 hand offs. 64 00:03:20,610 --> 00:03:24,650 And the most important thing that I want you to take on board is that it was simple. 65 00:03:24,770 --> 00:03:26,290 It's relatively lightweight. 66 00:03:26,330 --> 00:03:32,010 We we achieved something quite complex, quite advanced, and we did it all in a matter of a few minutes. 67 00:03:32,370 --> 00:03:36,690 Now, some exercises for you as of course, very important that you now come back and spend some time 68 00:03:36,690 --> 00:03:37,330 on this yourself. 69 00:03:37,330 --> 00:03:38,810 That is the best way to learn. 70 00:03:38,930 --> 00:03:43,650 So first of all, go through and identify the agentic design patterns that we used here. 71 00:03:44,090 --> 00:03:50,450 Now, there was a moment in this when we moved from from doing what anthropic would just describe as 72 00:03:50,450 --> 00:03:57,690 agent workflows to something that was really like agentic agents under their definition, although I 73 00:03:57,690 --> 00:03:59,650 think it's a little bit, a little bit hand-wavy. 74 00:03:59,650 --> 00:04:01,530 But still, there was a clear moment. 75 00:04:01,570 --> 00:04:02,530 Can you identify it? 76 00:04:02,570 --> 00:04:04,610 Can you spot what was the change? 77 00:04:04,650 --> 00:04:07,410 The small change that made that difference? 78 00:04:07,730 --> 00:04:10,290 And then try adding in more tools and more agents. 79 00:04:10,290 --> 00:04:11,490 This is a great way. 80 00:04:11,530 --> 00:04:12,970 This is so easy to extend. 81 00:04:12,970 --> 00:04:13,290 This. 82 00:04:13,330 --> 00:04:18,250 Obviously this is calling this an SDR was was maybe a bit of a stretch because it's really just a cold 83 00:04:18,250 --> 00:04:23,080 email writer, but it would be very easy to see how you could build this into something that's much 84 00:04:23,080 --> 00:04:29,640 more interactive and then a hard challenge for you, particularly if you've got some engineering experience. 85 00:04:29,920 --> 00:04:38,720 Figure out if you could turn this into something which is more of a longer living agent workflow, in 86 00:04:38,760 --> 00:04:45,840 that someone could reply to that email, and the agent would then be able to pick up and continue the 87 00:04:45,840 --> 00:04:46,640 conversation. 88 00:04:46,680 --> 00:04:51,040 Have an email based conversation about the company. 89 00:04:51,320 --> 00:04:56,160 Now the hard part about that is handling replies back from SendGrid that we'll need to do a little bit 90 00:04:56,160 --> 00:05:01,080 of research and understanding about things like webhooks and how that will work, and how you'd use 91 00:05:01,080 --> 00:05:06,240 that to trigger this processing, and how would you identify who you sent it to? 92 00:05:06,280 --> 00:05:11,680 So there's a fair amount of stuff in there, but it's engineering stuff rather than much genetic stuff. 93 00:05:11,680 --> 00:05:11,880 So. 94 00:05:11,880 --> 00:05:15,480 So you could be forgiven for skipping it if it doesn't appeal to you. 95 00:05:15,480 --> 00:05:20,680 But if it does, then, then that's a very interesting framework to put in place. 96 00:05:21,520 --> 00:05:26,160 And in terms of commercial applications, because you remember, I do always like to try and bring this 97 00:05:26,160 --> 00:05:28,870 back to how can you apply this to business? 98 00:05:29,110 --> 00:05:34,830 And hopefully I mean, obviously there's the fact that it could be literally a sales automation tool 99 00:05:34,830 --> 00:05:40,190 that could just be working to generate a bunch of different cold sales emails or email campaigns. 100 00:05:40,350 --> 00:05:46,030 And if you were to do this extra hard project, it could in fact be something that engages and continues 101 00:05:46,030 --> 00:05:47,070 the conversation. 102 00:05:47,070 --> 00:05:53,750 And generally, the point here is that you can really apply this to any end to end automation of these 103 00:05:53,750 --> 00:05:55,430 kinds of business processes. 104 00:05:55,430 --> 00:05:59,790 You can think of how you can apply this in this context of sales, but you can think of applying it 105 00:05:59,790 --> 00:06:01,310 to so many contexts. 106 00:06:01,350 --> 00:06:04,630 Imagine my day job at Nebula that involves recruitment. 107 00:06:04,910 --> 00:06:11,910 Imagine how applicable something like this is to that space as well, and many others. 108 00:06:12,070 --> 00:06:13,470 Think of a business area. 109 00:06:13,510 --> 00:06:20,510 Think of the kinds of activities that are done at scale with things like cold sales, outreach, and 110 00:06:20,510 --> 00:06:27,670 picture how this kind of approach agents collaborating with a level of autonomy and with use of tools 111 00:06:27,670 --> 00:06:32,710 and handoffs, can automate a complex business process.