WEBVTT 00:00:05.740 --> 00:00:09.220 You've probably heard the term AI agent a hundred 00:00:09.220 --> 00:00:12.080 times this year. Everyone's talking about it. 00:00:12.460 --> 00:00:14.900 But lately, a new phrase has started showing 00:00:14.900 --> 00:00:19.300 up. Multi -agent systems. Sounds technical, but 00:00:19.300 --> 00:00:22.420 it's actually pretty simple. A multi -agent system 00:00:22.420 --> 00:00:25.600 is just a group of AI agents that talk to each 00:00:25.600 --> 00:00:28.300 other, share tasks, and work together like a 00:00:28.300 --> 00:00:31.379 team. Instead of one agent doing everything alone, 00:00:32.030 --> 00:00:34.350 Multiple agents handle different parts of a job 00:00:34.350 --> 00:00:37.130 and pass the results between them. In today's 00:00:37.130 --> 00:00:39.369 episode, we're going to unpack what that means. 00:00:39.829 --> 00:00:42.109 I'll show you how multi -agent systems work, 00:00:42.270 --> 00:00:45.149 how they communicate, what can go wrong, and 00:00:45.149 --> 00:00:51.850 how you can use them to save real time in your 00:00:51.850 --> 00:00:54.310 daily work. We'll even walk through a live example 00:00:54.310 --> 00:00:56.869 of agents that research, write, and review content 00:00:56.869 --> 00:01:01.219 for you, all while you stay in control. By the 00:01:01.219 --> 00:01:03.840 end of this episode, you'll understand what separates 00:01:03.840 --> 00:01:07.620 basic automation from real collaboration between 00:01:07.620 --> 00:01:11.680 AI systems. Welcome back to Data and AI with 00:01:11.680 --> 00:01:14.819 Mukundan, where you learn AI by building. I'm 00:01:14.819 --> 00:01:17.120 Mukundan, and in this show, we talk about practical 00:01:17.120 --> 00:01:21.599 human ways to use AI tools, not for hype, but 00:01:21.599 --> 00:01:24.799 for real results. Now, in the last step... two 00:01:24.799 --> 00:01:27.900 episodes we talked about how single agents can 00:01:27.900 --> 00:01:31.579 plan your day or even manage your job hunt but 00:01:31.579 --> 00:01:33.760 in real life work doesn't happen in isolation 00:01:33.760 --> 00:01:36.680 you don't have one person who does everything 00:01:36.680 --> 00:01:40.239 you have a group of people who collaborate hand 00:01:40.239 --> 00:01:44.180 things off check each other's work and make sure 00:01:44.180 --> 00:01:47.859 the final result is strong and that's exactly 00:01:47.859 --> 00:01:52.519 what multi -agent systems are they are like your 00:01:52.519 --> 00:01:56.390 wife virtual team instead of one large model 00:01:56.390 --> 00:01:59.489 doing all the work you break the process into 00:01:59.489 --> 00:02:05.069 smaller career steps each agent gets a job each 00:02:05.069 --> 00:02:10.169 agent passes its results to the next one today 00:02:10.169 --> 00:02:13.030 i'm going to explain this in plain english and 00:02:13.030 --> 00:02:16.150 show you how to think about building your own 00:02:16.150 --> 00:02:21.370 ai team one that can work while you sleep So 00:02:21.370 --> 00:02:24.930 let's start with the basics. A single agent is 00:02:24.930 --> 00:02:29.250 like a solo worker. You give it a goal, it plans 00:02:29.250 --> 00:02:32.930 the steps, it acts and it learns from feedback. 00:02:33.669 --> 00:02:37.389 A multi -agent system is like a small team though. 00:02:37.650 --> 00:02:41.169 It's like a group of agents, each one good at 00:02:41.169 --> 00:02:44.289 a specific job. They can communicate with each 00:02:44.289 --> 00:02:47.830 other, divide the work, check each other's output 00:02:47.830 --> 00:02:52.500 and hand tasks back and forth. Now, let's say 00:02:52.500 --> 00:02:56.080 you wanted to write a research article. You could 00:02:56.080 --> 00:02:58.240 assign the first agent to gather information 00:02:58.240 --> 00:03:01.039 from trusted sources. The second agent could 00:03:01.039 --> 00:03:04.879 write the first draft. The third could edit it 00:03:04.879 --> 00:03:08.460 for clarity. The fourth could format it and get 00:03:08.460 --> 00:03:13.479 ready to post. So basically, each agent has a 00:03:13.479 --> 00:03:17.960 narrow focus. One plans, one writes, one reviews. 00:03:18.830 --> 00:03:21.409 They don't try to do everything at once. They 00:03:21.409 --> 00:03:25.610 just do one part well and pass the baton. So 00:03:25.610 --> 00:03:27.810 the real magic happens when they share context. 00:03:28.830 --> 00:03:31.530 That's what makes it feel like teamwork instead 00:03:31.530 --> 00:03:35.530 of just random automation. Here's how agents 00:03:35.530 --> 00:03:38.949 communicate. They don't talk like humans. They 00:03:38.949 --> 00:03:41.810 pass structured messages, small packets of information 00:03:41.810 --> 00:03:45.129 that tell the next agent what's done and what's 00:03:45.129 --> 00:03:49.060 next. Now imagine three agents working on a short 00:03:49.060 --> 00:03:52.900 report. The first one says, summarize the latest 00:03:52.900 --> 00:03:56.719 data on renewable energy. The second replies, 00:03:57.039 --> 00:03:59.819 summary complete. Here are the top five insights. 00:04:00.719 --> 00:04:04.000 The third looks at those insights, finds a gap 00:04:04.000 --> 00:04:07.199 and sends a note back. It says, double check 00:04:07.199 --> 00:04:11.319 insight number three. So the second one reviews 00:04:11.319 --> 00:04:13.180 it really and fixes the error and returns it. 00:04:13.930 --> 00:04:16.329 This back and forth is called message passing. 00:04:16.970 --> 00:04:20.029 It keeps everyone aligned. Now each agent knows 00:04:20.029 --> 00:04:22.889 what's been done, what's pending, and what to 00:04:22.889 --> 00:04:27.230 correct. That simple exchange is what allows 00:04:27.230 --> 00:04:30.589 multi -agent systems to coordinate just like 00:04:30.589 --> 00:04:34.310 human teams. There's accountability because each 00:04:34.310 --> 00:04:37.629 step is logged. There's reliability because if 00:04:37.629 --> 00:04:40.329 one fails, the others can pick up where it left 00:04:40.329 --> 00:04:44.360 off. And there's transparency. because you can 00:04:44.360 --> 00:04:48.540 always trace why something happened when the 00:04:48.540 --> 00:04:52.100 structure is missing agents start stepping on 00:04:52.100 --> 00:04:55.600 each other's toes when it's there the system 00:04:55.600 --> 00:04:59.540 feels stable and intelligent now let's take a 00:04:59.540 --> 00:05:02.899 real world example remember the job hunting agent 00:05:02.899 --> 00:05:06.180 we built in the last episode now let's make it 00:05:06.180 --> 00:05:09.519 a multi -agent here's how that would look in 00:05:09.519 --> 00:05:13.240 practice the first agent tracks to job applications 00:05:13.240 --> 00:05:16.500 every time you apply it adds the details the 00:05:16.500 --> 00:05:21.399 company role company the role date salary and 00:05:21.399 --> 00:05:26.139 link to your notion or google sheet the second 00:05:26.139 --> 00:05:29.439 agent does the research it looks up each company 00:05:29.439 --> 00:05:32.879 checks their glassdoor ratings recent funding 00:05:32.879 --> 00:05:36.779 and team size then it sends that data to the 00:05:36.779 --> 00:05:40.370 next agent The third agent prepares interview 00:05:40.370 --> 00:05:43.610 questions based on that information. It tailors 00:05:43.610 --> 00:05:47.209 them to the company and your resume. The fourth 00:05:47.209 --> 00:05:50.689 agent handles follow -ups. After the interview, 00:05:50.889 --> 00:05:53.889 it drafts a thank you email and sets a reminder 00:05:53.889 --> 00:05:58.329 to follow up if you don't hear back. Now each 00:05:58.329 --> 00:06:00.990 agent knows what the others are doing. When you 00:06:00.990 --> 00:06:03.709 apply to a new company, the tracker alerts the 00:06:03.709 --> 00:06:07.410 researcher. The researcher passes notes to the 00:06:07.410 --> 00:06:11.040 prep agent. The prep agent then alerts the follow 00:06:11.040 --> 00:06:14.199 -up agent once the interview is complete. The 00:06:14.199 --> 00:06:17.480 result is a system that behaves like an organized 00:06:17.480 --> 00:06:21.379 invisible team working behind the scenes. You 00:06:21.379 --> 00:06:23.680 still review everything, you're still the decision 00:06:23.680 --> 00:06:27.519 maker, but the heavy lifting is off of your plate. 00:06:30.439 --> 00:06:33.279 Now let's do another live demo where we have 00:06:33.279 --> 00:06:36.939 a simple multi -agent workflow. that can run 00:06:36.939 --> 00:06:40.899 on top of any LLM model. We'll build three agents, 00:06:41.060 --> 00:06:45.420 a researcher, a writer and a reviewer. The goal 00:06:45.420 --> 00:06:49.079 is to create a 200 -word LinkedIn post about 00:06:49.079 --> 00:06:54.360 three new AI tools that launched this week. The 00:06:54.360 --> 00:06:57.139 researcher starts by searching for the newest 00:06:57.139 --> 00:07:01.279 tools. It collects short descriptions and features. 00:07:02.659 --> 00:07:05.240 Writer then receives that input and drafts a 00:07:05.240 --> 00:07:08.899 clean, readable post summarizing the key takeaways. 00:07:10.120 --> 00:07:12.540 Reviewer checks that tone, corrects mistakes, 00:07:12.860 --> 00:07:16.439 and shortens the text where needed. You can watch 00:07:16.439 --> 00:07:19.639 them communicate. The researcher then sends also, 00:07:19.699 --> 00:07:23.019 here's what I found. Writer replies, draft complete. 00:07:24.379 --> 00:07:27.120 The reviewer comments, shorten paragraph 2, fix 00:07:27.120 --> 00:07:30.720 typos, rephrase, conclusion. Writer updates and 00:07:30.720 --> 00:07:33.399 resubmits. now within a minute you have a finished 00:07:33.399 --> 00:07:35.519 piece that's fact -checked clear and ready to 00:07:35.519 --> 00:07:40.540 post that's what a multi -functioning multi -agent 00:07:40.540 --> 00:07:43.819 system feels like something that's functioning 00:07:43.819 --> 00:07:47.459 at least well it doesn't rush right so it coordinates 00:07:47.459 --> 00:07:50.959 it's the digital version of a well -run project 00:07:50.959 --> 00:07:55.279 team of course things don't always go smoothly 00:07:55.279 --> 00:07:58.879 multi -agent systems fail when there's no hierarchy 00:07:59.759 --> 00:08:04.199 No memory or weak feedback loops. If every agent 00:08:04.199 --> 00:08:07.000 thinks it's in charge, they'll overwrite each 00:08:07.000 --> 00:08:10.120 other's work. If they forget what's already been 00:08:10.120 --> 00:08:13.720 done, they'll repeat the same task again. If 00:08:13.720 --> 00:08:16.839 there's no feedback, one small error multiplies 00:08:16.839 --> 00:08:20.439 as it moves through the chain. That's why guardrails 00:08:20.439 --> 00:08:24.319 matter. Always define who leads, who follows, 00:08:24.439 --> 00:08:27.920 and how feedback flows. The easiest way is to 00:08:27.920 --> 00:08:31.399 start with a single shared goal, two or three 00:08:31.399 --> 00:08:34.519 agents, and a rule for when humans get to approve. 00:08:35.940 --> 00:08:41.019 It's not about making machines work alone. It's 00:08:41.019 --> 00:08:43.820 about making them work with you, predictably 00:08:43.820 --> 00:08:48.620 and safely. Now, let's pause for a quick quiz 00:08:48.620 --> 00:08:52.179 to see if you're catching the pattern. I'll read 00:08:52.179 --> 00:08:54.899 five statements out loud. You can just say true 00:08:54.899 --> 00:08:59.600 or false in your head. One, multi -agent systems 00:08:59.600 --> 00:09:05.379 are just faster versions of single agents. I'll 00:09:05.379 --> 00:09:07.960 repeat that again. Multi -agent systems are just 00:09:07.960 --> 00:09:12.639 faster versions of single agents. And I'll just 00:09:12.639 --> 00:09:16.200 say the answer as well simultaneously. So the 00:09:16.200 --> 00:09:18.559 answer is false because multi -agent systems 00:09:18.559 --> 00:09:23.379 aren't faster, they're just smarter. Two, agents 00:09:23.379 --> 00:09:26.590 need... Clear communication to work together 00:09:26.590 --> 00:09:30.669 effectively Do is true communication is really 00:09:30.669 --> 00:09:34.149 everything Because the question is agents need 00:09:34.149 --> 00:09:36.049 clear communication to work together effectively 00:09:36.049 --> 00:09:41.490 and that is true three feedback loops only matter 00:09:41.490 --> 00:09:47.070 for solo agents Feedback loops only matter for 00:09:47.070 --> 00:09:50.350 solo agents and this is false because feedback 00:09:50.350 --> 00:09:53.330 loops matter even more when multiple agents are 00:09:53.330 --> 00:09:58.519 involved 4. Multi -agent systems work best when 00:09:58.519 --> 00:10:02.639 every agent has the same goal. 4 is false because 00:10:02.639 --> 00:10:05.200 they work best when each agent has a distinct 00:10:05.200 --> 00:10:08.639 role. Again, the question was multi -agent systems 00:10:08.639 --> 00:10:11.000 work best when every agent has the same goal. 00:10:12.659 --> 00:10:14.899 They work best when each agent has a distinct 00:10:14.899 --> 00:10:24.000 role. 5. Human approval steps prevent chaos. 00:10:24.970 --> 00:10:28.129 human approval steps prevent chaos five is true 00:10:28.129 --> 00:10:32.250 because the human in the loop checkpoints keep 00:10:32.250 --> 00:10:37.289 everything grounded so this one was true all 00:10:37.289 --> 00:10:39.389 right so now let's make this real for you here's 00:10:39.389 --> 00:10:42.090 your challenge for the week pick one workflow 00:10:42.090 --> 00:10:46.289 in your life that always takes too long something 00:10:46.289 --> 00:10:49.850 with at least two steps then imagine what it 00:10:49.850 --> 00:10:53.389 would like look like if two simple agents handled 00:10:54.000 --> 00:10:58.120 those steps and talk to each other. For example, 00:10:58.179 --> 00:11:00.519 if you have one agent summarize your meeting 00:11:00.519 --> 00:11:02.679 notes, then the second agent could turn those 00:11:02.679 --> 00:11:06.000 summaries into action items. You don't even have 00:11:06.000 --> 00:11:09.259 to build it yet. Just describe how they would 00:11:09.259 --> 00:11:12.620 pass information, what they would need to remember 00:11:12.620 --> 00:11:16.879 and where would you step in for approval. That's 00:11:16.879 --> 00:11:18.759 how you start thinking like a system designer. 00:11:19.960 --> 00:11:22.340 Once you can imagine how the pieces fit together, 00:11:22.919 --> 00:11:26.440 Building it is easy. Now, before we wrap up, 00:11:26.539 --> 00:11:30.100 I wanted to leave you with one thought. Automation 00:11:30.100 --> 00:11:35.299 saves you minutes. Collaboration saves you hours. 00:11:36.580 --> 00:11:39.220 Single agents are great at doing one task well. 00:11:40.159 --> 00:11:43.220 Multi -agent systems are powerful because they 00:11:43.220 --> 00:11:46.100 think ahead, hand off work, and close the loop. 00:11:46.960 --> 00:11:49.120 If this episode helped you understand how to 00:11:49.120 --> 00:11:52.039 design those systems, Share it with one teammate 00:11:52.039 --> 00:11:55.779 who's experimenting with AI workflows. Now next 00:11:55.779 --> 00:11:57.740 week, we'll go one level deeper into something 00:11:57.740 --> 00:12:01.559 new. How to combine human decisions with agent 00:12:01.559 --> 00:12:04.759 actions. It's about learning how to supervise 00:12:04.759 --> 00:12:07.679 and delegate to your AI like a real manager. 00:12:08.639 --> 00:12:12.480 Until then, remember this. The best AI systems 00:12:12.480 --> 00:12:16.019 don't just work for you. They work with you.