# You’re Not Behind (Yet): Learn AI Agents in 13 Minutes
Canonical: https://social-archive.org/jack_bauer111/IxzGKnayG6
Original URL: https://www.youtube.com/watch?v=P5sKKnWCvzk
Author: Sandeep Swadia
Platform: youtube
Share mode: full
## Content
Subscribe to my newsletter: https://sandeepswadia.beehiiv.com/ Take us on your morning run or commute, follow us on Spotify: https://open.spotify.com/show/033nSSACs3S4g5NUIrXnHx Most people still use AI like a better search box, but the real shift is AI agents, systems that decide the next action, not just the next word. I explain the difference between prompts and agents using ARR, show what’s happening “under the hood” with four roles, and map how agents adapt through an OODA loop when workflows break. I also cover why agents fail in real life: they amplify vague thinking and bad processes, so you need a GPS check before automating anything. The opportunity isn’t broad intelligence; it’s narrow, specific agents that solve repeated, hated tasks, as output gets cheap and judgment, taste, and standards become more valuable. --- ## Transcript [0:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=0s) Most people think they're using AI well [0:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=2s) when they get a decent answer from chat [0:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=4s) GPT. That was enough 6 months ago. It's [0:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=7s) not enough anymore. The next shift is AI [0:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=10s) agents, and [music] the gap between [0:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=12s) people who understand them and people [0:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=14s) who don't, it's about to get very [0:16](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=16s) expensive. I've spent years in the [0:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=18s) boardrooms of billion-dollar companies, [0:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=20s) and the good news here is that agents [0:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=23s) are much simpler than most people think. [0:26](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=26s) So, in this video, I'll show you exactly [0:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=28s) how they work, when to use them, and how [0:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=31s) to start before everyone else catches [0:33](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=33s) up. An AI prompt and an AI agent are [0:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=37s) completely different, but most of us are [0:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=39s) still stuck with old habits. Let me give [0:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=41s) you the simplest way to think about this [0:44](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=44s) before we go any further. This is our [0:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=46s) first framework, right off the bat. I [0:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=48s) call it ARR. If a task is autonomous, [0:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=51s) recurring, and reviewable, it's a strong [0:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=54s) candidate for an agent. If it needs live [0:57](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=57s) judgment, or it only happens once, or [1:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=60s) can't be reviewed clearly, then use a [1:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=63s) prompt. That one distinction alone will [1:05](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=65s) put you miles ahead of most people using [1:08](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=68s) AI today. The internet gave us search, [1:11](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=71s) so we started Googling. AI gave us LLMs, [1:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=74s) or large language models, but most [1:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=77s) people today still think of AI as a [1:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=80s) glorified search. Now, agents are here, [1:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=83s) and we're still making the same mistake [1:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=85s) again. We think of agents as just more [1:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=88s) capable chatbots. They're not. A chatbot [1:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=91s) waits for your next prompt. An agent [1:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=94s) figures out its next move. Prompting is [1:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=97s) like sitting next to a student driver. [1:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=99s) You still have to guide them, correct [1:42](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=102s) them, and stay very alert. An agent, on [1:44](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=104s) the other hand, is a hired driver. You [1:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=106s) set the destination, hand over the keys, [1:49](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=109s) and just sit in the backseat. It handles [1:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=111s) the route, the traffic, and all the [1:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=113s) step-by-step decisions. That's the [1:55](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=115s) mental shift we have to make. So, here's [1:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=118s) a prompt. Write me a LinkedIn post and [2:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=121s) here's an agent. Watch my [music] [2:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=122s) industry every Monday, find the three [2:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=124s) most relevant stories, study my previous [2:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=127s) post, draft a new post based on those [2:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=130s) stories in my voice, revise against my [2:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=132s) style, and schedule it for Tuesday [2:15](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=135s) morning. That's the power of AI agents. [2:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=138s) And to wield that power, you need to [2:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=140s) know what's actually running under the [2:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=143s) hood. Everyone's talking about AI [2:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=145s) agents. Almost nobody can tell you [2:27](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=147s) what's actually happening inside. A [2:29](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=149s) chatbot predicts the next word. [2:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=151s) >> [music] [2:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=151s) >> An agent decides the next action. Here's [2:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=154s) how a chatbot actually works. It's a [2:36](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=156s) large language model. When you type a [2:38](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=158s) question, it's going to break that [2:40](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=160s) question into small units of words [2:42](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=162s) called tokens and it converts them into [2:45](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=165s) numbers. And then it just finishes the [2:47](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=167s) sentence. So, if you said, "Jack fell [2:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=170s) down and broke his crown." Now, you know [2:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=173s) [music] that, but LLM does not know that [2:56](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=176s) rhyme. It will know words like bones and [2:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=178s) heart and crown. [music] [3:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=180s) And all of those words could make sense, [3:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=183s) but based on its training, it predicts [3:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=186s) that the most likely next word, given [3:08](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=188s) that line, is crown. It's based on [3:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=190s) probabilities. The agent has the same [3:13](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=193s) language model in the center, but now [3:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=197s) there are four workers around [music] [3:19](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=199s) it. Analyst, planner, operator, auditor. [3:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=203s) One finds the pattern, [3:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=205s) >> [music] [3:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=205s) >> one decides the plan, one does the work, [3:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=208s) one checks the results. Let's make this [3:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=210s) real. You can give an agent some [3:32](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=212s) instruction like, "Every Monday at 7:00 [3:35](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=215s) a.m., review the past week's customer [3:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=217s) support tickets, sales notes, and [3:40](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=220s) product feedback. Identify three biggest [3:44](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=224s) recurring issues, summarize what [3:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=226s) changed, and email my leadership team a [3:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=230s) one-page weekly brief." That's [music] [3:52](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=232s) it. I mean, those are lots of steps, but [3:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=234s) the agent will read the tickets, notes, [3:57](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=237s) and feedback, and find the pattern [3:59](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=239s) analyst. [music] It will then decide [4:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=241s) what matters most and what belongs in [4:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=243s) the brief planner. It will write and [4:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=246s) send the update operator. And then it'll [4:09](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=249s) check for weak logic, missing context, [4:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=252s) or sloppy conclusions, and it'll refine [4:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=254s) it auditor. Now, by Monday morning, the [4:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=257s) brief is in your team's inbox. You did [4:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=260s) not write the report, you did not [4:21](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=261s) analyze it, you just assigned the job of [4:24](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=264s) four people to one agent. So, here's [4:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=268s) your move. Tonight, open ChatGPT agent [4:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=271s) mode and give it one recurring task, [4:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=274s) >> [music] [4:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=274s) >> but then watch what it does. You'll see [4:36](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=276s) all four workers show up in real [music] [4:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=279s) time. That's the anatomy of an AI agent. [4:43](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=283s) Now that you understand the parts, let's [4:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=286s) look at the entire loop. The best thing [4:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=288s) about agents is that they can adapt when [4:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=291s) things go wrong. This is what makes [4:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=294s) agents genuinely different from [4:56](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=296s) everything that came before it. In the [4:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=298s) 1970s, there was an Air Force Colonel [5:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=301s) John Boyd, and he studied a very [5:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=304s) intriguing puzzle from the Korean War. [5:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=306s) American pilots in their F-86 kept [5:09](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=309s) beating technically superior Soviet MiG. [5:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=312s) Now, the MiG was faster and it could [5:15](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=315s) climb higher. It should have won, but it [5:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=318s) didn't. And Boyd eventually found the [5:21](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=321s) difference. The American pilots could [5:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=323s) see more from their cockpits, and they [5:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=325s) could adapt faster. So, they got inside [5:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=328s) the enemy's decision cycle before the [5:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=330s) enemy could respond. He called that loop [5:33](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=333s) the OODA loop: observe, orient, decide, [5:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=337s) act. [music] And in the world of agents, [5:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=339s) it's the same thing. That is the real [5:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=341s) test of an agent. When the obvious path [5:43](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=343s) fails, can it choose a better one? Can [5:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=346s) it go through its own UDA loop? So, let [5:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=348s) me give you a concrete example. You can [5:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=350s) build an automated workflow. Every [5:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=353s) Friday, check this [music] week's [5:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=354s) grocery prices, build my shopping list, [5:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=358s) and place the order. It [music] works [6:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=360s) every Friday. Until one week, your usual [6:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=363s) item is out of stock, [music] [6:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=364s) and you have six friends coming for [6:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=366s) dinner on Saturday. So, that automated [6:09](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=369s) workflow is going to break. Not because [6:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=372s) it's dumb, but because it's designed to [6:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=374s) be obedient. It's designed to not think [6:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=377s) on its own. An agent, on the other hand, [6:19](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=379s) does something very unique. It sees the [6:21](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=381s) usual list. It sees that it's not [6:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=383s) working. It finds substitutes. It [6:26](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=386s) adjusts quantities for six people. It [6:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=388s) checks your calendar, sees the dinner, [6:32](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=392s) and rebuilds the entire order around it. [6:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=394s) A workflow can follow the process. An [6:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=397s) agent can reroute it completely. [music] [6:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=399s) That's the difference. So, when someone [6:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=401s) says they built an agent, ask one [6:43](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=403s) question. When the first path breaks, [6:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=406s) does the agent keep following the [6:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=408s) script, or can it find [music] a better [6:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=410s) path? Can it find another way? That's [6:52](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=412s) the agent adapting in real time. So, [6:56](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=416s) this begs the question, right? If agents [6:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=418s) have autonomy, why do they still fail so [7:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=421s) often [7:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=422s) >> [music] [7:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=422s) >> in real life? [7:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=424s) That's next. The most dangerous thing [7:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=426s) about AI agents is that they will do [7:08](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=428s) wrong things faster and with more [7:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=430s) confidence than you ever could. An agent [7:13](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=433s) is not magic. It's a multiplier. I was [7:15](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=435s) working with the aboard and leadership [7:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=438s) team of large consumer company. Yeah, [7:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=440s) I'm still working with them. And they're [7:22](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=442s) profitable, well-run, great CEO. And [7:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=445s) when I asked what was stopping them from [7:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=448s) using AI to drive customer acquisition, [7:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=450s) for example, the CMO responded, "We have [7:35](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=455s) all the data, but we'll still need to [7:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=457s) build a clean process, so we can turn [7:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=461s) that into something useful, something [7:43](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=463s) insightful. And I asked where the real [7:46](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=466s) challenge was and she said, "You know, [7:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=468s) we need the right people in the seats [7:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=470s) first." That's the story everywhere. [7:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=473s) Most AI problems are human problems in [7:56](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=476s) disguise. An agent is just a mirror. It [7:59](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=479s) reflects the quality of your thinking [8:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=481s) back at you. It just amplifies it. Give [8:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=483s) an agent vague goals, sloppy directions, [8:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=487s) and no way to get feedback and it will [8:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=490s) drive the car straight into the tree [8:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=492s) faster and with more confidence than you [8:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=494s) ever could. Here's the dangerous part. [8:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=497s) An agent doesn't fix bad thinking, it [8:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=500s) formalizes it. Usually the agent fails [8:22](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=502s) because the human was vague, not because [8:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=505s) the underlying model was bad or [8:27](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=507s) anything. So, before you automate [8:29](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=509s) anything, run a GPS check. Goal, proof, [8:33](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=513s) steps. Goal, can I define the goal in [8:36](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=516s) one sentence very clearly? [music] [8:38](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=518s) Proof, can I tell what good looks like [8:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=521s) and how will I know if the agent got it [8:44](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=524s) right? And steps, can I describe each [8:47](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=527s) and every step very clearly without a [8:49](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=529s) lot of hand-waving? Unless you can do [8:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=531s) those three things very well, your agent [8:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=533s) is not going to make any difference. [8:55](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=535s) I'll give [music] you an example. Here [8:57](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=537s) are two instructions for your agent. [8:59](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=539s) First one, "Summarize my emails every [9:01](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=541s) morning." It's good. And here's the [9:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=542s) second one, "Every morning at 7:00 a.m., [9:05](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=545s) read my unread emails, categorize them [9:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=547s) by urgency, draft replies to routine [9:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=550s) messages, and flag anything from my top [9:13](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=553s) five customers." So you see there is a [9:15](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=555s) difference between those two [9:16](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=556s) instructions and that [music] gap is [9:19](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=559s) exactly where the mess lives. The [9:21](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=561s) winners who can wield the power of AI [9:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=563s) agents aren't just [music] going to be [9:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=565s) engineers. They'll be the people who [9:27](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=567s) understand their work deeply enough to [9:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=570s) define it precisely. [music] [9:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=571s) Most companies want AI everywhere. The [9:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=574s) ones actually winning are obsessively [9:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=577s) narrow. If clarity is a bottleneck, then [9:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=579s) the opportunity is not broad [9:41](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=581s) intelligence, it's narrow ownership. I [9:44](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=584s) was visiting the customer conference of [9:47](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=587s) construction software company that I [9:49](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=589s) work with, and the product lead was on [9:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=591s) stage showing a demo of a single agent [9:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=594s) >> [music] [9:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=594s) >> that was focused on a very specific [9:56](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=596s) problem, collecting field data for a [9:59](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=599s) specific type of customers in a specific [10:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=602s) type of situation. And it was a beta [10:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=604s) launch. The demo worked mostly with a [10:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=607s) few minor glitches here and there, but [10:09](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=609s) when he showed the QR code at the end, [10:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=612s) every hand in the conference room went [10:14](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=614s) up with their phones. Everyone took a [10:16](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=616s) picture [music] [10:17](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=617s) because it solved a very specific, but [10:20](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=620s) very real pain they all had been living [10:23](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=623s) with for decades. [music] That is where [10:25](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=625s) the real opportunity is, in your career [10:28](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=628s) or in your company. Narrow [music] [10:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=630s) focus. Here's the test. Find a highly [10:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=634s) specific task [music] [10:36](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=636s) people hate doing, but they have to do [10:38](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=638s) it repeatedly. That's where the money [10:40](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=640s) is. You know, we're entering an age [10:42](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=642s) where we'll have more agents than human [10:45](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=645s) beings on this planet. On the business [10:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=648s) side, for every software company that [10:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=650s) exists today, there will be an agent [10:53](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=653s) company trying to dethrone it. The [10:55](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=655s) winners won't build the broadest agents [10:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=658s) first. They'll build the one that [11:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=660s) understands one workflow, [music] [11:03](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=663s) one market, and one kind of user pain [11:06](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=666s) better than everyone else. By the way, [11:08](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=668s) if you want to keep this conversation [11:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=670s) going, I write a short newsletter once a [11:13](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=673s) week, just useful ideas, tools, honest [11:16](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=676s) reflections. You can subscribe below. [11:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=678s) It's free. AI will reshape almost every [11:22](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=682s) role, but it won't replace what makes [11:24](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=684s) you irreplaceable. You know, AI is a [11:27](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=687s) giant decoupling machine. For most of [11:30](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=690s) our modern history, your income was tied [11:33](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=693s) to your hours. [music] [11:35](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=695s) Even at the top, you're always trading [11:37](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=697s) time for decisions. Agents are breaking [11:40](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=700s) that link for the first time. Now they [11:43](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=703s) do the work and you scale your judgment [11:45](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=705s) in areas where it matters most. And that [11:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=708s) changes what's valuable. We're entering [11:51](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=711s) an era of infinite output, content, [11:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=714s) code, and analysis all becoming super [11:57](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=717s) cheap. When intelligence becomes that [12:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=720s) cheap, judgment becomes even more [12:02](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=722s) expensive. When output becomes infinite, [12:05](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=725s) taste becomes scarce. Every time you [12:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=727s) define a task clearly enough for an [12:10](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=730s) agent to run it, you're not just [12:12](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=732s) training the system, you're clarifying [12:15](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=735s) your own standards. You're learning what [12:18](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=738s) good actually looks like. Sure, some [12:21](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=741s) roles will be reshaped. The paralegal, [12:24](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=744s) the junior analyst, the coordinator. But [12:27](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=747s) every disruption has always created new [12:29](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=749s) roles that never existed before. [music] [12:31](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=751s) Before the internet, nobody imagined the [12:34](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=754s) role of an online community manager. The [12:36](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=756s) question is not whether the shift [12:39](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=759s) happens. It is whether you shape it or [12:42](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=762s) get shaped by it. The most valuable [12:45](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=765s) person is no longer the one who can [12:48](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=768s) [music] think the fastest. It's the one [12:50](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=770s) who can define good work, spot bad work, [12:54](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=774s) and know when to trust an agent and when [12:58](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=778s) to trust a human. That is where your [13:00](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=780s) value is going to move. Ironically, AI [13:04](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=784s) will make human life less robotic. [13:07](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=787s) Thank you. [13:08](https://www.youtube.com/watch?v=P5sKKnWCvzk&t=788s) And I love you.
