WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:23.000 examine the ideas shaping the future of artificial minds. 5 00:00:23.839 --> 00:00:27.320 Welcome back to the deep Dive. Today's Tuesday, February seventeenth, 6 00:00:27.399 --> 00:00:30.679 twenty twenty six, and honestly, staring at that date on 7 00:00:30.719 --> 00:00:33.399 the calendar, I'm having a bit of a where did 8 00:00:33.399 --> 00:00:36.119 the time go? Moment? Right, But it's more than that, 9 00:00:36.200 --> 00:00:40.479 it's a how did the world change this fast? Moment? 10 00:00:40.640 --> 00:00:44.759 It's disorienting, isn't it. The velocity of the last twenty 11 00:00:44.799 --> 00:00:47.520 four months specifically has been unlike anything we've seen in 12 00:00:47.560 --> 00:00:48.159 tech history. 13 00:00:48.399 --> 00:00:50.439 I mean really, I was actually digging through my digital 14 00:00:50.560 --> 00:00:53.079 archives this morning, you know, looking at some old notes 15 00:00:53.119 --> 00:00:56.399 from late twenty twenty two, early twenty twenty three. Do 16 00:00:56.399 --> 00:00:59.320 you remember the initial chat GPT craze vividly? Oh? 17 00:00:59.399 --> 00:01:03.159 Absolutely, the wow factor of a computer writing a haikup 18 00:01:03.320 --> 00:01:04.159 or a little poem. 19 00:01:04.400 --> 00:01:06.560 Right, we were all losing our minds because a machine 20 00:01:06.560 --> 00:01:09.239 could summarize an email or write a wedding toast. It 21 00:01:09.239 --> 00:01:10.439 felt like absolute magic. 22 00:01:10.599 --> 00:01:11.000 It did. 23 00:01:11.159 --> 00:01:13.000 But looking back from where we sit right now in 24 00:01:13.040 --> 00:01:15.959 early twenty twenty six, it feels a little bit like 25 00:01:15.959 --> 00:01:17.359 looking at a telegraph machine. 26 00:01:17.480 --> 00:01:20.959 It does. It seems almost quaint because those systems, as 27 00:01:20.959 --> 00:01:23.959 impressive as they were, they were fundamentally passive. 28 00:01:24.000 --> 00:01:25.359 I'm passive. That's the word. 29 00:01:25.680 --> 00:01:27.719 They were waiting for you. You are the pilot. They 30 00:01:27.719 --> 00:01:30.280 were just the engine. If you stop typing, they stopped 31 00:01:30.280 --> 00:01:31.439 thinking exactly. 32 00:01:31.480 --> 00:01:34.040 You had to prompt engineer your way to a result, 33 00:01:34.280 --> 00:01:39.920 step by step by step. But today, today we are 34 00:01:39.959 --> 00:01:43.319 talking about the shift that makes that era look like 35 00:01:43.400 --> 00:01:46.840 the Stone Age. We're talking about the realization that AI 36 00:01:47.560 --> 00:01:50.319 isn't just a tool you talk to anymore. It's a 37 00:01:50.319 --> 00:01:54.920 tool that does things. It is moved from chat to act. 38 00:01:55.200 --> 00:01:57.920 That is the critical distinction. That is the whole ballgame. 39 00:01:58.159 --> 00:02:01.159 We have left the era conversation and entered the era 40 00:02:01.319 --> 00:02:02.079 of agency. 41 00:02:02.400 --> 00:02:05.959 So today we're doing a deep dive into agentic AI. 42 00:02:06.480 --> 00:02:08.240 And I want to be clear right off the bat, 43 00:02:08.439 --> 00:02:10.520 this isn't just another buzzword we're throwing around. 44 00:02:10.599 --> 00:02:11.520 No, it's very real. 45 00:02:11.719 --> 00:02:14.199 If you've been following the industry analysis, you know that 46 00:02:14.280 --> 00:02:18.520 Gardner officially declared twenty twenty six the year of agentic AI. 47 00:02:18.479 --> 00:02:20.800 And that's a significant declaration. You know, they don't just 48 00:02:20.840 --> 00:02:22.360 throw those titles around lightly. 49 00:02:22.520 --> 00:02:24.759 Well, I have to play Devil's advocate for a second here, 50 00:02:25.039 --> 00:02:27.960 Gardner says a lot of things. In twenty twenty three, 51 00:02:28.280 --> 00:02:31.199 they were hyping the metaverse pretty hard, and we all 52 00:02:31.280 --> 00:02:32.360 saw how that played out. 53 00:02:32.439 --> 00:02:34.319 That's a fair point, a very fair point. 54 00:02:34.599 --> 00:02:38.199 Is this just another marketing slogan to sell enterprise software 55 00:02:38.319 --> 00:02:40.919 or is there actually something different happening in the silicon. 56 00:02:40.639 --> 00:02:42.960 It is fundamentally different. And the proof isn't just in 57 00:02:43.000 --> 00:02:45.199 the marketing. It's in the deployment numbers. It's in the 58 00:02:45.240 --> 00:02:48.280 actual usage. The statistic that jumped out at me from 59 00:02:48.319 --> 00:02:50.560 the research stack you sent over is that by the 60 00:02:50.639 --> 00:02:52.199 end of this year or so by the end of 61 00:02:52.199 --> 00:02:56.240 twenty twenty six, forty percent of all enterprise applications will 62 00:02:56.280 --> 00:02:59.000 embed task specific AI agents. 63 00:02:59.039 --> 00:02:59.520 Forty percent. 64 00:02:59.560 --> 00:03:03.080 That can be forty four zero. Now, to put that 65 00:03:03.120 --> 00:03:06.319 in perspective, twelve months ago that number was less than 66 00:03:06.400 --> 00:03:07.000 five percent. 67 00:03:07.159 --> 00:03:07.599 Wow. 68 00:03:07.919 --> 00:03:10.080 That is not a trend line. That is a vertical wall. 69 00:03:10.159 --> 00:03:11.199 It's a step change. 70 00:03:11.319 --> 00:03:14.520 That's adoption at a speed that usually breaks things. That's 71 00:03:14.560 --> 00:03:15.680 scary fast. 72 00:03:15.439 --> 00:03:17.400 It is, and the money follows. I mean, the market 73 00:03:17.439 --> 00:03:19.800 is surging from about what seven point eight billion dollars 74 00:03:19.840 --> 00:03:22.439 sort of projected fifty two billion dollars by twenty thirty. 75 00:03:22.800 --> 00:03:25.800 But honestly, forget the dollars for a second. The reason 76 00:03:25.840 --> 00:03:29.800 the spending is skyrocketing is utility. The promise of generative 77 00:03:29.840 --> 00:03:32.280 AI was I'll help you, right. The promise of a 78 00:03:32.360 --> 00:03:34.719 gentic AI is I'll do the work for you. 79 00:03:35.000 --> 00:03:36.960 Okay, So that is our mission today. We are going 80 00:03:37.039 --> 00:03:39.759 to cut through the noise. I want to unpack what 81 00:03:39.800 --> 00:03:43.400 it actually means to have a digital colleague. And I 82 00:03:43.400 --> 00:03:44.919 don't want the brochure version. 83 00:03:44.800 --> 00:03:45.960 Right, Let's get under the hood. 84 00:03:46.080 --> 00:03:48.840 I want to know how these things work. What is 85 00:03:48.879 --> 00:03:51.879 physically happening when I give an instruction, Why is this 86 00:03:51.919 --> 00:03:54.639 happening now in early twenty twenty six and not two 87 00:03:54.759 --> 00:03:57.159 years ago? And frankly, I want to talk about the 88 00:03:57.199 --> 00:03:59.199 messiness of it all, because I don't buy that it's 89 00:03:59.240 --> 00:03:59.960 all smooth saling. 90 00:04:00.680 --> 00:04:03.800 It is definitely not all smooth sailing, but it is 91 00:04:03.879 --> 00:04:06.960 the tipping point where the capability finally caught up to 92 00:04:07.000 --> 00:04:07.759 the science fiction. 93 00:04:08.120 --> 00:04:10.080 So let's start with the basics. But let's go deep 94 00:04:10.759 --> 00:04:15.360 definitions agentic AI. It sounds academical, little jargoning. 95 00:04:15.120 --> 00:04:15.479 I'd be it. 96 00:04:15.560 --> 00:04:17.600 Yeah, if I'm explaining this to a smart friend who 97 00:04:17.639 --> 00:04:19.720 fell into a coma in twenty twenty four and just 98 00:04:19.759 --> 00:04:22.839 woke up, what is the core difference between the AI 99 00:04:22.920 --> 00:04:25.839 they knew and the agentic AI of twenty twenty six. 100 00:04:26.160 --> 00:04:28.519 The simplest way to frame it is reactive. 101 00:04:28.160 --> 00:04:30.879 Versus proactive, reactive versus proactive. 102 00:04:30.519 --> 00:04:34.480 Traditional generative AI. The stuff from the coma days was reactive. 103 00:04:34.920 --> 00:04:36.800 If you asked a question, it predicts the next word. 104 00:04:36.839 --> 00:04:38.720 It gives you an answer. Maybe it writes some code, 105 00:04:39.160 --> 00:04:41.519 but then it stops. It goes dormant. It has no 106 00:04:41.759 --> 00:04:42.600 drive of its own. 107 00:04:43.040 --> 00:04:46.399 It's a super smart encyclopedia, but it doesn't care if 108 00:04:46.399 --> 00:04:48.199 the job gets done. It has no skin in the 109 00:04:48.199 --> 00:04:49.040 game exactly. 110 00:04:49.079 --> 00:04:52.279 It has no agency. Agentic AI, on the other hand, 111 00:04:52.399 --> 00:04:54.920 has a goal. You don't give it a prompt. You 112 00:04:54.959 --> 00:04:55.959 give it an objective. 113 00:04:56.160 --> 00:05:00.360 An objective, so not write me an email about but 114 00:05:00.519 --> 00:05:01.519 something bigger. 115 00:05:01.319 --> 00:05:03.839 Much bigger, you say, book me a flight to London 116 00:05:03.920 --> 00:05:07.319 under eight hundred dollars, or refactor this entire code base 117 00:05:07.399 --> 00:05:11.040 to be Python three compliant. Or plan this whole marketing campaign, 118 00:05:11.319 --> 00:05:13.240 and the agent figures out the how. 119 00:05:13.399 --> 00:05:16.240 That's the key, right, the how. It's not just filling 120 00:05:16.240 --> 00:05:17.120 in the blanks anymore. 121 00:05:17.240 --> 00:05:20.240 Yes, it breaks the high level goal down into steps. 122 00:05:20.399 --> 00:05:23.079 It executes those steps. It checks its own work, and 123 00:05:23.240 --> 00:05:25.920 this is the most important part. It keeps going until 124 00:05:25.920 --> 00:05:27.920 the job is done or it hits a hard wall 125 00:05:27.959 --> 00:05:28.600 and needs help. 126 00:05:28.720 --> 00:05:30.879 That checking its own work part is where I want 127 00:05:30.920 --> 00:05:32.600 to drill in. Because in the old days, if I 128 00:05:32.639 --> 00:05:34.920 asked a model to write code, it would write it. 129 00:05:35.040 --> 00:05:37.959 If the code was broken, too bad, yep, your problem. 130 00:05:38.000 --> 00:05:40.160 Now I had to copy and paste the error message 131 00:05:40.160 --> 00:05:41.680 back in and say, hey, you messed up. 132 00:05:41.879 --> 00:05:43.199 You were the feedback loop. 133 00:05:43.199 --> 00:05:45.879 You were part of the machine, right, I was the debugger. 134 00:05:46.199 --> 00:05:49.480 In an agentic system, the softer is the feedback loop. 135 00:05:50.240 --> 00:05:52.600 This brings us to the anatomy of how these things function. 136 00:05:53.639 --> 00:05:56.639 It's a pattern that's often referred to as the REACT. 137 00:05:56.240 --> 00:05:59.560 Pattern react ore eACT yes, short. 138 00:05:59.319 --> 00:06:03.360 For reason and act. It's a continuous loop that well, 139 00:06:03.800 --> 00:06:05.800 it mimics how a human solves a problem. 140 00:06:05.879 --> 00:06:07.920 Okay, let's make this tangible. I want to walk through 141 00:06:07.920 --> 00:06:09.759 this loop. Let's say I am an agent I'm a 142 00:06:09.759 --> 00:06:13.000 personal assistant agent. You give me a goal, find out 143 00:06:13.040 --> 00:06:15.439 if my favorite band is playing in Chicago this year 144 00:06:15.439 --> 00:06:16.160 and get tickets. 145 00:06:16.399 --> 00:06:17.480 Okay, good example. 146 00:06:17.680 --> 00:06:20.879 In the old version, the LM just hallucinates a date. 147 00:06:20.959 --> 00:06:23.600 Because it's training data is two years old. It makes 148 00:06:23.600 --> 00:06:24.439 something up right. 149 00:06:24.519 --> 00:06:27.000 It confidently tells you, yes, they are playing in twenty 150 00:06:27.040 --> 00:06:29.040 twenty one, and you're just frustrated. 151 00:06:29.199 --> 00:06:31.879 So how does the react loop handle this? What's the 152 00:06:31.959 --> 00:06:33.560 first thing I the agent do. 153 00:06:33.959 --> 00:06:37.160 The first step is perception. The agent has to gather 154 00:06:37.240 --> 00:06:40.360 information from its environment. It's not looking at its internal 155 00:06:40.360 --> 00:06:43.600 training data anymore. It needs fresh data, live data. 156 00:06:43.680 --> 00:06:45.360 Well knows it doesn't know precisely. 157 00:06:45.600 --> 00:06:47.920 The reasoning part of its brain says, I don't know 158 00:06:48.000 --> 00:06:50.000 the answer to this. I need to use a tool 159 00:06:50.079 --> 00:06:50.959 to find out. 160 00:06:51.040 --> 00:06:53.000 A tool, not just its own memory. 161 00:06:53.120 --> 00:06:57.040 Yes, this is crucial. The agent has access to tools. 162 00:06:57.120 --> 00:06:59.319 Think of these as APIs. It has a web search tool, 163 00:06:59.399 --> 00:07:01.480 it has a ticket tool, it has access to your 164 00:07:01.519 --> 00:07:05.560 calendar tool. So step one, it perceives the request. 165 00:07:05.759 --> 00:07:06.120 Okay. 166 00:07:06.279 --> 00:07:09.920 Step two planning. It engages in what we call chain 167 00:07:10.000 --> 00:07:13.360 of thought reasoning. It literally talks to itself in the background. 168 00:07:13.360 --> 00:07:14.920 It forms a strategy. 169 00:07:14.560 --> 00:07:16.639 What is it saying, like? What does that internal monologue 170 00:07:16.639 --> 00:07:16.959 look like? 171 00:07:17.079 --> 00:07:20.720 It's saying, goal, buy tickets for the user's favorite band 172 00:07:20.759 --> 00:07:23.680 in Chicago. Step one, I need to know who their 173 00:07:23.720 --> 00:07:25.839 favorite band is. Let me check my memory. Ah, it's 174 00:07:25.879 --> 00:07:29.040 the Lumineers. Okay. Step two search for the Lumineers tour 175 00:07:29.079 --> 00:07:33.399 dates Chicago twenty twenty six. Step three, check users calendar 176 00:07:33.439 --> 00:07:36.759 for conflicts on those dates. Step four if free, search 177 00:07:36.800 --> 00:07:40.120 for tickets under their price limit. Step five purchase. 178 00:07:40.439 --> 00:07:42.680 It creates a checklist, an actual plan. 179 00:07:42.519 --> 00:07:45.519 Of attack exactly, and then comes the action. It calls 180 00:07:45.560 --> 00:07:47.639 the web search tool. It executes that query. 181 00:07:47.879 --> 00:07:51.639 Okay, so I the agent run the search, and let's 182 00:07:51.639 --> 00:07:54.399 say the search comes back and says no tour dates 183 00:07:54.439 --> 00:07:57.279 listed for Chicago in twenty twenty four of the bot 184 00:07:57.319 --> 00:07:59.560 would just say sorry, no dates. 185 00:07:59.519 --> 00:08:02.240 And give up right end of story. But an agent 186 00:08:02.319 --> 00:08:06.120 moose to the next phase, observation and adaptation. It looks 187 00:08:06.120 --> 00:08:09.680 at the output of the action result no dates in Chicago. 188 00:08:10.639 --> 00:08:13.680 It evaluates this against the goal. The goal has not 189 00:08:13.800 --> 00:08:15.920 been met, so it iterates. 190 00:08:15.680 --> 00:08:18.360 It tries a different angle, doesn't just quit, yes. 191 00:08:18.720 --> 00:08:21.680 It updates the plan. The reasoning part kicks in again. 192 00:08:21.959 --> 00:08:24.360 It might think, Okay, maybe they aren't playing Chicago proper, 193 00:08:24.399 --> 00:08:26.480 but what about a nearby city. Let me expand my 194 00:08:26.519 --> 00:08:29.360 search radius to one hundred miles. Or maybe the dates 195 00:08:29.439 --> 00:08:31.319 just haven't been announced yet. Let me check the band's 196 00:08:31.360 --> 00:08:34.159 official Twitter account instead of a general search. That's a 197 00:08:34.200 --> 00:08:34.840 better source. 198 00:08:35.000 --> 00:08:37.360 It pivots. Yeah, it problem solves. 199 00:08:37.320 --> 00:08:42.440 It pivots, It runs the loop again, search Twitter, observe result. Ah, 200 00:08:42.720 --> 00:08:45.120 I see they're playing a festival in Milwaukee, which is 201 00:08:45.200 --> 00:08:47.840 ninety minutes away. New plan. Check if the user is 202 00:08:47.840 --> 00:08:50.960 willing to travel, cross reference their travel history. Hmm, they've 203 00:08:51.000 --> 00:08:54.200 gone to Milwaukee for concerts before. This is a viable option. 204 00:08:54.399 --> 00:08:57.120 That is a fundamental shift. It feels much more like 205 00:08:57.159 --> 00:08:59.080 how I work. Yeah, if I hit a roadblock, I 206 00:08:59.080 --> 00:09:00.960 don't just power down. I look for a window. I 207 00:09:00.960 --> 00:09:02.240 try something else exactly. 208 00:09:02.279 --> 00:09:06.559 That persistence, that goal seeking behavior is what defines agency. 209 00:09:07.039 --> 00:09:09.000 But I want to clarify the architecture here because I 210 00:09:09.000 --> 00:09:12.440 think a lot of people, even tech people, still confuse 211 00:09:12.519 --> 00:09:15.320 the agent with the LLM. They think GPT five is 212 00:09:15.360 --> 00:09:15.720 the agent. 213 00:09:15.879 --> 00:09:17.840 No, and that's a vital distinction to make. Think of 214 00:09:17.879 --> 00:09:20.759 the LLM, the GGT five or the clawed opus model 215 00:09:20.840 --> 00:09:25.399 as the brain. The brain, it provides the reasoning, the language, understanding, 216 00:09:25.440 --> 00:09:28.519 the logic. But a brain in a jar cannot type 217 00:09:28.559 --> 00:09:31.080 an email. A brain in a jar cannot access a 218 00:09:31.159 --> 00:09:32.519 database or click a button. 219 00:09:32.639 --> 00:09:35.600 It needs a body, It needs hands, It needs a body. 220 00:09:36.159 --> 00:09:39.360 The agent is the scaffolding around the brain. It includes 221 00:09:39.360 --> 00:09:41.440 the tools we mentioned, which are the arms and legs. 222 00:09:41.480 --> 00:09:45.120 It includes the sensors for perception, and crucially, it includes 223 00:09:45.559 --> 00:09:46.600 memory systems. 224 00:09:47.039 --> 00:09:48.840 This is something that came up a lot, the importance 225 00:09:48.840 --> 00:09:52.399 of memory because llms typically have amnesia. Right, yeah, you 226 00:09:52.399 --> 00:09:54.559 start a new chat. It has no idea who you are. 227 00:09:54.679 --> 00:09:57.120 They do every time you start a new chat. It's 228 00:09:57.120 --> 00:10:01.039 a blank slate. But an agent needs contact. It needs 229 00:10:01.120 --> 00:10:03.960 to know what you asked last week. It needs to 230 00:10:04.000 --> 00:10:06.120 know your preferences, your constraints. 231 00:10:06.240 --> 00:10:08.600 So how do they solve the amnesia problem? How does 232 00:10:08.600 --> 00:10:09.240 it remember? 233 00:10:09.519 --> 00:10:11.360 We use something called vector databases. 234 00:10:11.440 --> 00:10:14.039 Okay, without getting two into the map, weeds visualize that 235 00:10:14.080 --> 00:10:14.799 for me. What is that? 236 00:10:15.240 --> 00:10:18.799 Imagine a standard database, like a spreadsheet. You search for 237 00:10:18.840 --> 00:10:21.279 the term client X, and it finds the row that 238 00:10:21.320 --> 00:10:24.879 says client X. It's a literal keyword match. Simple a 239 00:10:24.960 --> 00:10:27.759 vector database is more like a three D map of concepts. 240 00:10:28.240 --> 00:10:30.840 When the agent learns something like the user hates early 241 00:10:30.879 --> 00:10:34.039 morning flights, it converts that concept into a string of 242 00:10:34.120 --> 00:10:37.240 numbers a vector, and places it in a specific spot 243 00:10:37.279 --> 00:10:39.440 on this three D map. That spot might be near 244 00:10:39.559 --> 00:10:44.320 other concepts like travel preferences or sleep habits, or being grumpy. 245 00:10:44.600 --> 00:10:47.759 So it's storing the meaning, not just the keywords exactly. 246 00:10:48.320 --> 00:10:50.879 So three weeks later, when you say book a trip 247 00:10:50.919 --> 00:10:54.759 to Denver, the agent stands its memory map, sees the 248 00:10:54.799 --> 00:10:57.279 trip concept, and it looks at what other concepts are 249 00:10:57.320 --> 00:10:59.759 nearby in the map, and it finds the vector for 250 00:11:00.080 --> 00:11:03.639 hates early morning flights. It retrieves that memory and injects 251 00:11:03.679 --> 00:11:05.240 it into the planning step. 252 00:11:05.519 --> 00:11:07.679 So it doesn't just book the six point zero am 253 00:11:07.720 --> 00:11:11.240 flight because it's cheapest. It remembers all be cranky and 254 00:11:11.279 --> 00:11:13.000 filters that option out precisely. 255 00:11:13.159 --> 00:11:15.840 That long term persistence, that memory is what makes it 256 00:11:15.879 --> 00:11:17.639 a colleague, not just a calculator. 257 00:11:17.759 --> 00:11:20.240 So why is this happening now? I remember, back in 258 00:11:20.320 --> 00:11:23.320 early twenty twenty three, there was that project called autogpt. 259 00:11:23.480 --> 00:11:24.039 Do you remember that? 260 00:11:24.120 --> 00:11:26.120 Oh? Absolutely, everyone was talking about it. 261 00:11:26.120 --> 00:11:28.440 It was the GitHub repo heard round the world. Everyone 262 00:11:28.480 --> 00:11:32.039 installed it. It promised to do exactly this, run loops, 263 00:11:32.159 --> 00:11:35.919 achieve goals, but let's be honest, it kind of sucked. 264 00:11:36.039 --> 00:11:39.559 It was a fascinating failure, a glorious, very expensive failure. 265 00:11:39.600 --> 00:11:40.360 For many people. 266 00:11:40.639 --> 00:11:43.320 It would get stuck in these infinite loops. It would 267 00:11:43.519 --> 00:11:47.039 try to Google something, fail, try again, fail, and just 268 00:11:47.279 --> 00:11:50.279 burn through your API credits until you went broke. 269 00:11:50.519 --> 00:11:54.360 It was ambitious, but it was premature. In twenty twenty three, 270 00:11:54.600 --> 00:11:58.679 the underlying models, the brains, just weren't smart enough. They 271 00:11:58.759 --> 00:12:02.600 lacked the sophisticated reasoning capability to recover from errors or 272 00:12:02.639 --> 00:12:03.799 to pivot effectively. 273 00:12:04.039 --> 00:12:06.519 So what changed? What were the key breakthroughs in the 274 00:12:06.600 --> 00:12:10.039 last say, eighteen months that made twenty twenty six the 275 00:12:10.120 --> 00:12:10.840 year of the Agent. 276 00:12:11.279 --> 00:12:14.759 There were three main pillars that all came together. First, 277 00:12:14.879 --> 00:12:17.639 the reasoning models themselves got a lot better. Okay, we 278 00:12:17.679 --> 00:12:21.080 saw the release of open ais O one series and Thropics, 279 00:12:21.080 --> 00:12:25.600 Claude thinking models, and Google's Gemini Advanced. These models were 280 00:12:25.600 --> 00:12:29.559 trained specifically to think before they speak. They are significantly 281 00:12:29.600 --> 00:12:31.759 better at that planning step of the React loop. 282 00:12:31.840 --> 00:12:33.919 They make better checklist, more robust. 283 00:12:33.639 --> 00:12:36.279 Checklist, much better checklists, and they're less likely to be 284 00:12:36.320 --> 00:12:39.840 distracted or go off the rails. Second is context windows. 285 00:12:39.559 --> 00:12:42.200 Right, the amount of information they can hold it once exactly. 286 00:12:42.240 --> 00:12:45.080 We went from a few thousand tokens to models that 287 00:12:45.080 --> 00:12:47.960 can hold millions of tokens in their working memory. They 288 00:12:47.960 --> 00:12:50.799 can read an entire book or a massive code based 289 00:12:50.879 --> 00:12:53.279 instantly and understand it all at once, which. 290 00:12:53.039 --> 00:12:55.360 Gives them the situational awareness they need to make good 291 00:12:55.399 --> 00:12:56.200 decisions right. 292 00:12:56.519 --> 00:12:59.120 And the third, and this is geeky, but crucial, is 293 00:12:59.399 --> 00:13:00.919 reliable tool calling. 294 00:13:01.320 --> 00:13:02.000 What does that mean? 295 00:13:02.399 --> 00:13:04.639 In twenty twenty three, if you asked a model to 296 00:13:04.720 --> 00:13:07.639 use a calculator, it might get the syntax wrong. It 297 00:13:07.720 --> 00:13:10.240 might forget a bracket or a comma. It was clumsy. 298 00:13:10.879 --> 00:13:13.960 Today the models are fine tuned to interface with software 299 00:13:14.039 --> 00:13:17.240 with APIs almost perfectly. They know how to format the 300 00:13:17.279 --> 00:13:18.879 request correctly every time. 301 00:13:19.120 --> 00:13:21.000 It's the difference between a toddler trying to use a 302 00:13:21.000 --> 00:13:24.720 hammer and a master carpenter using a hammer. The intent 303 00:13:24.840 --> 00:13:27.279 was there, but the execution was messy. 304 00:13:27.440 --> 00:13:30.440 That's a perfect analogy, and that reliability is what allowed 305 00:13:30.559 --> 00:13:33.879 enterprise adoption to jump from five percent to forty percent. 306 00:13:34.159 --> 00:13:36.320 You couldn't put the toddler in charge of your payroll. 307 00:13:36.360 --> 00:13:38.200 You can put the carpenter in charge. 308 00:13:37.919 --> 00:13:40.559 Which brings us to the next big evolution. In this story, 309 00:13:40.919 --> 00:13:44.200 we've talked about the single agent, the single carpenter, but 310 00:13:44.279 --> 00:13:48.519 the industry seems to have realized that one superbot isn't 311 00:13:48.559 --> 00:13:51.159 the answer. Right, We're hearing a lot about the multi 312 00:13:51.200 --> 00:13:51.879 agent revolution. 313 00:13:52.120 --> 00:13:54.600 That's right. We moved from the solo model to the 314 00:13:54.639 --> 00:13:58.799 squad model, from a single agent to a team the squad. 315 00:13:59.480 --> 00:14:03.360 I like that, but play the skeptic for me. Why 316 00:14:03.399 --> 00:14:05.919 do I need a squad? If GPT six or whatever 317 00:14:05.919 --> 00:14:07.679 it is so smart, why can't it just do it all? 318 00:14:07.960 --> 00:14:11.080 It comes down to what researchers call the generalist curse. 319 00:14:11.399 --> 00:14:13.039 M HM, I unpack that for me. 320 00:14:13.200 --> 00:14:16.320 Imagine you have a brilliant human genius, just an Einstein 321 00:14:16.399 --> 00:14:18.840 level intellect. You ask them to write a beautiful poem. 322 00:14:18.840 --> 00:14:20.879 They write a great poem. Then you ask them to 323 00:14:20.879 --> 00:14:23.720 write a poem, and solve a complex calculus problem and 324 00:14:23.799 --> 00:14:26.159 debug a Python script all at the same time. 325 00:14:26.240 --> 00:14:28.000 Their brain melts. They'd be terrible at all. 326 00:14:28.000 --> 00:14:33.639 Three their performance in each individual task drops significantly. This 327 00:14:33.720 --> 00:14:37.360 happens to LLLMS two. It's a phenomenon called attention drift. 328 00:14:37.919 --> 00:14:40.399 When you ask one model to hold the context for 329 00:14:40.519 --> 00:14:45.480 five different disciplines legal coding, creative writing, data analysis, it 330 00:14:45.519 --> 00:14:48.440 gets mediocre at all of them. It can't specialize. 331 00:14:48.440 --> 00:14:51.159 So specialization is the fix. Just like with people. 332 00:14:51.360 --> 00:14:54.840 Specialization is the fix. Just like a real company. You 333 00:14:54.840 --> 00:14:57.480 don't hire one person to be your lawyer, your coder, 334 00:14:57.519 --> 00:15:00.679 and your graphic designer. You hire a team of SAR specialists. 335 00:15:00.960 --> 00:15:03.759 So walk me through how a squad tackles a problem. 336 00:15:03.840 --> 00:15:06.000 Let's say we're building a new software feature. How does 337 00:15:06.039 --> 00:15:06.360 that work? 338 00:15:06.480 --> 00:15:09.399 Okay, so you have a supervisor agent. This is the boss, 339 00:15:09.399 --> 00:15:12.360 the project manager. It takes your high level request build 340 00:15:12.360 --> 00:15:14.360 a user authentication page for a new app. 341 00:15:14.399 --> 00:15:16.519 And it doesn't write a single line of code itself, 342 00:15:16.720 --> 00:15:17.320 not a line. 343 00:15:17.440 --> 00:15:20.519 Its only job is to break the goal down and delegate. 344 00:15:20.080 --> 00:15:21.480 Who does it call? Who's on the team. 345 00:15:21.879 --> 00:15:24.639 It calls the researcher agent to look up best practices 346 00:15:24.679 --> 00:15:28.279 for secure logins. It calls the coder agent, which is 347 00:15:28.320 --> 00:15:32.200 fine tuned on code, to write the actual HTML and CSS. 348 00:15:32.759 --> 00:15:35.120 But here is the most critical edition. It calls the 349 00:15:35.120 --> 00:15:35.840 critic agent. 350 00:15:36.000 --> 00:15:37.519 The critic. I like the sound of that. 351 00:15:37.600 --> 00:15:40.000 The critic's only job is to look at the code 352 00:15:40.039 --> 00:15:42.799 the coder wrote and try to break it. To be adversarial. 353 00:15:42.879 --> 00:15:46.159 It finds bugs, It checks for security flaws, It says 354 00:15:46.200 --> 00:15:48.799 this code is garbage. It doesn't handle this edge case. 355 00:15:49.080 --> 00:15:52.679 Do it again. That is fascinating. We are literally building 356 00:15:52.840 --> 00:15:56.440 the creative tension and quality control of a human team 357 00:15:56.600 --> 00:15:57.919 into the software itself. 358 00:15:58.039 --> 00:16:00.960 We are we are replicating the address serial nature of 359 00:16:01.039 --> 00:16:05.080 human collaboration, and the data shows it works. Google Research 360 00:16:05.120 --> 00:16:08.440 published a paper just last month in January twenty twenty 361 00:16:08.440 --> 00:16:11.279 six that highlighted this specific dynamic. 362 00:16:11.360 --> 00:16:12.320 What was the key finding? 363 00:16:12.360 --> 00:16:15.039 They found that simply throwing more agents at a problem 364 00:16:15.080 --> 00:16:17.399 isn't always better. If you have one hundred agents all 365 00:16:17.440 --> 00:16:19.480 shouting at each other in a flat structure, you just 366 00:16:19.559 --> 00:16:20.279 get chaos. 367 00:16:20.440 --> 00:16:21.320 Right, too many cooks. 368 00:16:21.480 --> 00:16:25.399 Exactly. The key is smart orchestration. It's about the hierarchy. 369 00:16:26.039 --> 00:16:29.399 A small, well orchestrated team with a clear supervisor and 370 00:16:29.440 --> 00:16:33.879 critic beats a massive, disorganized mob every single time. 371 00:16:34.080 --> 00:16:36.519 So the manager agent is actually the most valuable piece 372 00:16:36.559 --> 00:16:38.919 of code in the stack. Its ability to lead is 373 00:16:38.960 --> 00:16:40.000 the real intelligence. 374 00:16:40.519 --> 00:16:45.600 Exactly. The ability to delegate, review, and synthesize is the bottleneck. 375 00:16:45.960 --> 00:16:47.320 Just like in a human organization. 376 00:16:47.519 --> 00:16:50.039 Who are the players driving this right now? If I 377 00:16:50.080 --> 00:16:52.360 want to hire a squad today, where do I go? 378 00:16:52.840 --> 00:16:55.320 It's the big names you'd expect, but the tools are new. 379 00:16:55.679 --> 00:16:59.960 Open Ai has their Frontier platform and the Codex desktop app, 380 00:17:00.120 --> 00:17:04.319 which is heavily focused on this orchestration layer. Anthropic just 381 00:17:04.400 --> 00:17:08.440 released Claude Opus four point six, which specifically features agent 382 00:17:08.519 --> 00:17:11.640 teams that you can spin up with basically one click. 383 00:17:11.839 --> 00:17:13.119 You just define the roles. 384 00:17:13.240 --> 00:17:15.359 And for the developers listening who want to build their. 385 00:17:15.279 --> 00:17:19.119 Own, you have open source frameworks like langgraf and crew ai. 386 00:17:19.599 --> 00:17:22.279 These are libraries that let you build these custom org 387 00:17:22.359 --> 00:17:25.920 charts in code. You define agent A is the researcher, 388 00:17:26.000 --> 00:17:28.680 agent B is the writer, and you draw the lines 389 00:17:28.720 --> 00:17:30.440 for how information flows between them. 390 00:17:30.559 --> 00:17:33.319 It's like building a circuit board, but the components are 391 00:17:33.359 --> 00:17:34.279 intelligent workers. 392 00:17:34.480 --> 00:17:36.759 That's a beautiful way to put it. It's exactly that. 393 00:17:37.079 --> 00:17:39.680 Okay, we've been very abstract for a while. I want 394 00:17:39.720 --> 00:17:41.519 to ground this. I want to know what this feels 395 00:17:41.519 --> 00:17:43.440 like on it Tuesday morning at my desk. 396 00:17:43.559 --> 00:17:44.039 Let's do it. 397 00:17:44.079 --> 00:17:47.160 The briefing material had a scenario about a Q two 398 00:17:47.359 --> 00:17:51.160 sales trip. I want to roleplay this minute by minute, 399 00:17:51.880 --> 00:17:54.240 because I think anyone who has ever planned a multi 400 00:17:54.240 --> 00:17:58.400 stop business trip knows the specific kind of administrative hell 401 00:17:58.440 --> 00:17:59.079