You’re Not Behind (Yet): Learn AI Agents in 13 Minutes — Transcript

Full transcript

1. 0:00Most people think they're using AI well
2. 0:02when they get a decent answer from chat
3. 0:04GPT. That was enough 6 months ago. It's
4. 0:07not enough anymore. The next shift is AI
5. 0:10agents, and [music] the gap between
6. 0:12people who understand them and people
7. 0:14who don't, it's about to get very
8. 0:16expensive. I've spent years in the
9. 0:18boardrooms of billion-dollar companies,
10. 0:20and the good news here is that agents
11. 0:23are much simpler than most people think.
12. 0:26So, in this video, I'll show you exactly
13. 0:28how they work, when to use them, and how
14. 0:31to start before everyone else catches
15. 0:33up. An AI prompt and an AI agent are
16. 0:37completely different, but most of us are
17. 0:39still stuck with old habits. Let me give
18. 0:41you the simplest way to think about this
19. 0:44before we go any further. This is our
20. 0:46first framework, right off the bat. I
21. 0:48call it ARR. If a task is autonomous,
22. 0:51recurring, and reviewable, it's a strong
23. 0:54candidate for an agent. If it needs live
24. 0:57judgment, or it only happens once, or
25. 1:00can't be reviewed clearly, then use a
26. 1:03prompt. That one distinction alone will
27. 1:05put you miles ahead of most people using
28. 1:08AI today. The internet gave us search,
29. 1:11so we started Googling. AI gave us LLMs,
30. 1:14or large language models, but most
31. 1:17people today still think of AI as a
32. 1:20glorified search. Now, agents are here,
33. 1:23and we're still making the same mistake
34. 1:25again. We think of agents as just more
35. 1:28capable chatbots. They're not. A chatbot
36. 1:31waits for your next prompt. An agent
37. 1:34figures out its next move. Prompting is
38. 1:37like sitting next to a student driver.
39. 1:39You still have to guide them, correct
40. 1:42them, and stay very alert. An agent, on
41. 1:44the other hand, is a hired driver. You
42. 1:46set the destination, hand over the keys,
43. 1:49and just sit in the backseat. It handles
44. 1:51the route, the traffic, and all the
45. 1:53step-by-step decisions. That's the
46. 1:55mental shift we have to make. So, here's
47. 1:58a prompt. Write me a LinkedIn post and
48. 2:01here's an agent. Watch my [music]
49. 2:02industry every Monday, find the three
50. 2:04most relevant stories, study my previous
51. 2:07post, draft a new post based on those
52. 2:10stories in my voice, revise against my
53. 2:12style, and schedule it for Tuesday
54. 2:15morning. That's the power of AI agents.
55. 2:18And to wield that power, you need to
56. 2:20know what's actually running under the
57. 2:23hood. Everyone's talking about AI
58. 2:25agents. Almost nobody can tell you
59. 2:27what's actually happening inside. A
60. 2:29chatbot predicts the next word.
61. 2:31>> [music]
62. 2:31>> An agent decides the next action. Here's
63. 2:34how a chatbot actually works. It's a
64. 2:36large language model. When you type a
65. 2:38question, it's going to break that
66. 2:40question into small units of words
67. 2:42called tokens and it converts them into
68. 2:45numbers. And then it just finishes the
69. 2:47sentence. So, if you said, "Jack fell
70. 2:50down and broke his crown." Now, you know
71. 2:53[music] that, but LLM does not know that
72. 2:56rhyme. It will know words like bones and
73. 2:58heart and crown. [music]
74. 3:00And all of those words could make sense,
75. 3:03but based on its training, it predicts
76. 3:06that the most likely next word, given
77. 3:08that line, is crown. It's based on
78. 3:10probabilities. The agent has the same
79. 3:13language model in the center, but now
80. 3:17there are four workers around [music]
81. 3:19it. Analyst, planner, operator, auditor.
82. 3:23One finds the pattern,
83. 3:25>> [music]
84. 3:25>> one decides the plan, one does the work,
85. 3:28one checks the results. Let's make this
86. 3:30real. You can give an agent some
87. 3:32instruction like, "Every Monday at 7:00
88. 3:35a.m., review the past week's customer
89. 3:37support tickets, sales notes, and
90. 3:40product feedback. Identify three biggest
91. 3:44recurring issues, summarize what
92. 3:46changed, and email my leadership team a
93. 3:50one-page weekly brief." That's [music]
94. 3:52it. I mean, those are lots of steps, but
95. 3:54the agent will read the tickets, notes,
96. 3:57and feedback, and find the pattern
97. 3:59analyst. [music] It will then decide
98. 4:01what matters most and what belongs in
99. 4:03the brief planner. It will write and
100. 4:06send the update operator. And then it'll
101. 4:09check for weak logic, missing context,
102. 4:12or sloppy conclusions, and it'll refine
103. 4:14it auditor. Now, by Monday morning, the
104. 4:17brief is in your team's inbox. You did
105. 4:20not write the report, you did not
106. 4:21analyze it, you just assigned the job of
107. 4:24four people to one agent. So, here's
108. 4:28your move. Tonight, open ChatGPT agent
109. 4:31mode and give it one recurring task,
110. 4:34>> [music]
111. 4:34>> but then watch what it does. You'll see
112. 4:36all four workers show up in real [music]
113. 4:39time. That's the anatomy of an AI agent.
114. 4:43Now that you understand the parts, let's
115. 4:46look at the entire loop. The best thing
116. 4:48about agents is that they can adapt when
117. 4:51things go wrong. This is what makes
118. 4:54agents genuinely different from
119. 4:56everything that came before it. In the
120. 4:581970s, there was an Air Force Colonel
121. 5:01John Boyd, and he studied a very
122. 5:04intriguing puzzle from the Korean War.
123. 5:06American pilots in their F-86 kept
124. 5:09beating technically superior Soviet MiG.
125. 5:12Now, the MiG was faster and it could
126. 5:15climb higher. It should have won, but it
127. 5:18didn't. And Boyd eventually found the
128. 5:21difference. The American pilots could
129. 5:23see more from their cockpits, and they
130. 5:25could adapt faster. So, they got inside
131. 5:28the enemy's decision cycle before the
132. 5:30enemy could respond. He called that loop
133. 5:33the OODA loop: observe, orient, decide,
134. 5:37act. [music] And in the world of agents,
135. 5:39it's the same thing. That is the real
136. 5:41test of an agent. When the obvious path
137. 5:43fails, can it choose a better one? Can
138. 5:46it go through its own UDA loop? So, let
139. 5:48me give you a concrete example. You can
140. 5:50build an automated workflow. Every
141. 5:53Friday, check this [music] week's
142. 5:54grocery prices, build my shopping list,
143. 5:58and place the order. It [music] works
144. 6:00every Friday. Until one week, your usual
145. 6:03item is out of stock, [music]
146. 6:04and you have six friends coming for
147. 6:06dinner on Saturday. So, that automated
148. 6:09workflow is going to break. Not because
149. 6:12it's dumb, but because it's designed to
150. 6:14be obedient. It's designed to not think
151. 6:17on its own. An agent, on the other hand,
152. 6:19does something very unique. It sees the
153. 6:21usual list. It sees that it's not
154. 6:23working. It finds substitutes. It
155. 6:26adjusts quantities for six people. It
156. 6:28checks your calendar, sees the dinner,
157. 6:32and rebuilds the entire order around it.
158. 6:34A workflow can follow the process. An
159. 6:37agent can reroute it completely. [music]
160. 6:39That's the difference. So, when someone
161. 6:41says they built an agent, ask one
162. 6:43question. When the first path breaks,
163. 6:46does the agent keep following the
164. 6:48script, or can it find [music] a better
165. 6:50path? Can it find another way? That's
166. 6:52the agent adapting in real time. So,
167. 6:56this begs the question, right? If agents
168. 6:58have autonomy, why do they still fail so
169. 7:01often
170. 7:02>> [music]
171. 7:02>> in real life?
172. 7:04That's next. The most dangerous thing
173. 7:06about AI agents is that they will do
174. 7:08wrong things faster and with more
175. 7:10confidence than you ever could. An agent
176. 7:13is not magic. It's a multiplier. I was
177. 7:15working with the aboard and leadership
178. 7:18team of large consumer company. Yeah,
179. 7:20I'm still working with them. And they're
180. 7:22profitable, well-run, great CEO. And
181. 7:25when I asked what was stopping them from
182. 7:28using AI to drive customer acquisition,
183. 7:30for example, the CMO responded, "We have
184. 7:35all the data, but we'll still need to
185. 7:37build a clean process, so we can turn
186. 7:41that into something useful, something
187. 7:43insightful. And I asked where the real
188. 7:46challenge was and she said, "You know,
189. 7:48we need the right people in the seats
190. 7:50first." That's the story everywhere.
191. 7:53Most AI problems are human problems in
192. 7:56disguise. An agent is just a mirror. It
193. 7:59reflects the quality of your thinking
194. 8:01back at you. It just amplifies it. Give
195. 8:03an agent vague goals, sloppy directions,
196. 8:07and no way to get feedback and it will
197. 8:10drive the car straight into the tree
198. 8:12faster and with more confidence than you
199. 8:14ever could. Here's the dangerous part.
200. 8:17An agent doesn't fix bad thinking, it
201. 8:20formalizes it. Usually the agent fails
202. 8:22because the human was vague, not because
203. 8:25the underlying model was bad or
204. 8:27anything. So, before you automate
205. 8:29anything, run a GPS check. Goal, proof,
206. 8:33steps. Goal, can I define the goal in
207. 8:36one sentence very clearly? [music]
208. 8:38Proof, can I tell what good looks like
209. 8:41and how will I know if the agent got it
210. 8:44right? And steps, can I describe each
211. 8:47and every step very clearly without a
212. 8:49lot of hand-waving? Unless you can do
213. 8:51those three things very well, your agent
214. 8:53is not going to make any difference.
215. 8:55I'll give [music] you an example. Here
216. 8:57are two instructions for your agent.
217. 8:59First one, "Summarize my emails every
218. 9:01morning." It's good. And here's the
219. 9:02second one, "Every morning at 7:00 a.m.,
220. 9:05read my unread emails, categorize them
221. 9:07by urgency, draft replies to routine
222. 9:10messages, and flag anything from my top
223. 9:13five customers." So you see there is a
224. 9:15difference between those two
225. 9:16instructions and that [music] gap is
226. 9:19exactly where the mess lives. The
227. 9:21winners who can wield the power of AI
228. 9:23agents aren't just [music] going to be
229. 9:25engineers. They'll be the people who
230. 9:27understand their work deeply enough to
231. 9:30define it precisely. [music]
232. 9:31Most companies want AI everywhere. The
233. 9:34ones actually winning are obsessively
234. 9:37narrow. If clarity is a bottleneck, then
235. 9:39the opportunity is not broad
236. 9:41intelligence, it's narrow ownership. I
237. 9:44was visiting the customer conference of
238. 9:47construction software company that I
239. 9:49work with, and the product lead was on
240. 9:51stage showing a demo of a single agent
241. 9:54>> [music]
242. 9:54>> that was focused on a very specific
243. 9:56problem, collecting field data for a
244. 9:59specific type of customers in a specific
245. 10:02type of situation. And it was a beta
246. 10:04launch. The demo worked mostly with a
247. 10:07few minor glitches here and there, but
248. 10:09when he showed the QR code at the end,
249. 10:12every hand in the conference room went
250. 10:14up with their phones. Everyone took a
251. 10:16picture [music]
252. 10:17because it solved a very specific, but
253. 10:20very real pain they all had been living
254. 10:23with for decades. [music] That is where
255. 10:25the real opportunity is, in your career
256. 10:28or in your company. Narrow [music]
257. 10:30focus. Here's the test. Find a highly
258. 10:34specific task [music]
259. 10:36people hate doing, but they have to do
260. 10:38it repeatedly. That's where the money
261. 10:40is. You know, we're entering an age
262. 10:42where we'll have more agents than human
263. 10:45beings on this planet. On the business
264. 10:48side, for every software company that
265. 10:50exists today, there will be an agent
266. 10:53company trying to dethrone it. The
267. 10:55winners won't build the broadest agents
268. 10:58first. They'll build the one that
269. 11:00understands one workflow, [music]
270. 11:03one market, and one kind of user pain
271. 11:06better than everyone else. By the way,
272. 11:08if you want to keep this conversation
273. 11:10going, I write a short newsletter once a
274. 11:13week, just useful ideas, tools, honest
275. 11:16reflections. You can subscribe below.
276. 11:18It's free. AI will reshape almost every
277. 11:22role, but it won't replace what makes
278. 11:24you irreplaceable. You know, AI is a
279. 11:27giant decoupling machine. For most of
280. 11:30our modern history, your income was tied
281. 11:33to your hours. [music]
282. 11:35Even at the top, you're always trading
283. 11:37time for decisions. Agents are breaking
284. 11:40that link for the first time. Now they
285. 11:43do the work and you scale your judgment
286. 11:45in areas where it matters most. And that
287. 11:48changes what's valuable. We're entering
288. 11:51an era of infinite output, content,
289. 11:54code, and analysis all becoming super
290. 11:57cheap. When intelligence becomes that
291. 12:00cheap, judgment becomes even more
292. 12:02expensive. When output becomes infinite,
293. 12:05taste becomes scarce. Every time you
294. 12:07define a task clearly enough for an
295. 12:10agent to run it, you're not just
296. 12:12training the system, you're clarifying
297. 12:15your own standards. You're learning what
298. 12:18good actually looks like. Sure, some
299. 12:21roles will be reshaped. The paralegal,
300. 12:24the junior analyst, the coordinator. But
301. 12:27every disruption has always created new
302. 12:29roles that never existed before. [music]
303. 12:31Before the internet, nobody imagined the
304. 12:34role of an online community manager. The
305. 12:36question is not whether the shift
306. 12:39happens. It is whether you shape it or
307. 12:42get shaped by it. The most valuable
308. 12:45person is no longer the one who can
309. 12:48[music] think the fastest. It's the one
310. 12:50who can define good work, spot bad work,
311. 12:54and know when to trust an agent and when
312. 12:58to trust a human. That is where your
313. 13:00value is going to move. Ironically, AI
314. 13:04will make human life less robotic.
315. 13:07Thank you.
316. 13:08And I love you.