TOP 2% Engineering: /PLAN 2026 — Transcript

Full transcript

1. 0:00What's up engineers? Andy Dev Dan here.
2. 0:02I [music] would place myself in the top
3. 0:0410% of engineers using agentic
4. 0:07technology. Where would you place
5. 0:09yourself? [music] Do you think you're
6. 0:10top 15, top 10, top five? I want to
7. 0:12become a top 2% [music] engineer and I
8. 0:15want to take you along with me. In fact,
9. 0:17I want you to surpass [music]
10. 0:19what I can do. The best way to uplevel
11. 0:22what you can do is to create a plan.
12. 0:24[music] And that's exactly what I've
13. 0:26done. Let me share with you my roadmap
14. 0:28for becoming a top 2% agentic engineer
15. 0:32in 2026. This is everything I'm thinking
16. 0:35about. This is everything that I believe
17. 0:37based on all the conversations, all the
18. 0:39work I've done over the past years. This
19. 0:41is what I think it looks like to become
20. 0:43a top 2% a Gentic engineer in 2026.
21. 0:48Let's first start with a couple big
22. 0:50announcements that are going to tie in
23. 0:52really well. So, first off, Gemini 3
24. 0:54Flash is absolutely insane. This model
25. 0:57should not exist. Why do I say that? Why
26. 1:00should this model not exist? It's
27. 1:01because it is a top three intelligence
28. 1:04model. Top five price. This hasn't
29. 1:07happened before, by the way. You don't
30. 1:08get intelligence and price. On top of
31. 1:11that, what else do we have? We have the
32. 1:13top speed. Some measurements are coming
33. 1:15in at 200 tokens per second. It is a
34. 1:17lightning fast model. Gemini 3 flash
35. 1:20inputs text, images, audio, video, and
36. 1:23PDF. And just real quick here, you can
37. 1:25see this in the usage numbers, right?
38. 1:27You have to ignore all the benchmarks,
39. 1:29all these auxiliary numbers at some
40. 1:31point and just look at the raw usage
41. 1:32data. If we search for Gemini 3 flash
42. 1:35preview, check this out on Open Router,
43. 1:37385% token increase usage. It's up in
44. 1:4013th place. I'm predicting this goes up
45. 1:42to the top 10, potentially even the top
46. 1:44five by the end of the year. So, it's a
47. 1:47fantastic model. I've been using this
48. 1:48thing alongside Opus 4.5. So, you know,
49. 1:52what do these releases mean for
50. 1:54engineers? Let's set up for 2026, right?
51. 1:57What do these releases mean? What's
52. 1:58coming next for us? All right, we have
53. 2:00Gemini 3 Flash, Opus 4.5, and GBG 5.2.
54. 2:05These are all stateofthe-art correct
55. 2:07models. There has never been more
56. 2:09opportunity for us engineers. We're
57. 2:11going to break down that opportunity for
58. 2:132026 right now in this video. These
59. 2:15releases are telling us there are no
60. 2:17winners. And this is a great thing,
61. 2:19right? Competition drives innovation and
62. 2:22it also drives price down. That means
63. 2:25intelligence per token is going up. And
64. 2:28that means you and I, the engineer, we
65. 2:30can do more. Take a look at these
66. 2:32prices. Gemini 2.5 flash 30 cents.
67. 2:34Gemini 3 flash 50 cents. A lot of
68. 2:36engineers see a price increase here. I
69. 2:38see a price decrease. Why is that? It's
70. 2:42because Gemini 3 Flash is competing with
71. 2:45top state-of-the-art models. Okay,
72. 2:48intelligence per token has gone up. That
73. 2:50means cost per intelligence or cost per
74. 2:54token has gone down. You get more for
75. 2:56less. That's a big theme. Keep that in
76. 2:59mind as we work through this. All right,
77. 3:01the model ecosystem is competitive. So,
78. 3:03this is great. Like I mentioned, this
79. 3:04gives us optionality. Out of all these
80. 3:06things, there is one that is the most
81. 3:09important. What do these releases mean
82. 3:11for us engineers?
83. 3:13Models are no longer the limitation.
84. 3:16Tricky question for you. Do you know
85. 3:18what the limitation is? And this brings
86. 3:20us to the central thesis for my plan for
87. 3:232026 for top 2% engineering. If you want
88. 3:27to be the very best or at least approach
89. 3:29the best or improve, this is what I
90. 3:32think you need. Andrew Carpathy calls
91. 3:34the next 10 years the year of agents. I
92. 3:36think that's right. But we need to dial
93. 3:38in what is 2026? What is the theme for
94. 3:412026? I believe this is the year of
95. 3:44trust. What do I mean by that exactly?
96. 3:47Every one of these bets, every
97. 3:48prediction, every action I think top
98. 3:50engineers will be taking comes down to
99. 3:52one question. Do you trust your agents?
100. 3:56This is the year of trust. Do you trust
101. 4:00your agents? What do I mean by this?
102. 4:02There are powerful agentic systems,
103. 4:06agentic workflows, AI developer
104. 4:08workflows that you can build that
105. 4:10increases the amount of work you can
106. 4:12hand off to your agents. The question is
107. 4:15always this. Why haven't you pushed your
108. 4:17agents to do more? Why haven't you built
109. 4:19that next system? What's stopping you
110. 4:21from automating that task you did
111. 4:23yesterday? Right? And I think all of it,
112. 4:26and I mean all of it, boils down to one
113. 4:28word. You and I, we do not trust our
114. 4:31agents. And more specifically, this is a
115. 4:34gray question. It's not a yes or no
116. 4:37question. The true question here is how
117. 4:39much do you trust your agents? So, this
118. 4:42is why we are framing everything as the
119. 4:44year of trust. Let's pause for a second.
120. 4:46Why do we care about trust? We care
121. 4:48about trust because trust equals speed.
122. 4:51Speed equals iteration and iteration
123. 4:54equals impact. Think about a team you've
124. 4:57worked with. Think about a technology
125. 4:59you've used, right? I love the example
126. 5:01claw code running Opus 4.5. I trust that
127. 5:05model and that agent harness. Why?
128. 5:07Because I know it's going to deliver.
129. 5:09This gives me concrete speed. That speed
130. 5:11gives me iteration on the problems I'm
131. 5:14solving. And what does that give me? The
132. 5:16more iterations, the more cycles that
133. 5:18you get, that directly increases your
134. 5:21impact. All right. So why do we care
135. 5:23about trust? It's because it gives us
136. 5:25impact. So, this is why I think this
137. 5:27central thesis is very powerful for
138. 5:30thinking about how to become a top 2%
139. 5:32engineer in 2026. This is everything
140. 5:35we're going to discuss here. I'm just
141. 5:36going to lay it all out for you. If
142. 5:38you're interested in these ideas, stick
143. 5:40around. If you believe and you're on
144. 5:42board with my trust thesis, like this is
145. 5:45what it's all about. All right? Top 2%
146. 5:46engineers are going to make these bets
147. 5:48to effectively increase their trust. If
148. 5:51you don't like these ideas, if you're
149. 5:52not interested, do me a quick favor.
150. 5:54Copy that URL. uh for this video and
151. 5:57send it to your competition real quick.
152. 5:59Okay, [laughter]
153. 6:00so before we jump right in, let me just
154. 6:02say, you know, I do this every single
155. 6:05year. All right, in 2023, I made 24
156. 6:08bets, 16 were right. 2024, I made 41
157. 6:11bets. Tons of bets there. 36 were right.
158. 6:14All right, this year, as you'll see if
159. 6:16you stick around, I made 15 bets, 13
160. 6:18were right. Accuracy is going up. And
161. 6:20this year, I'm going to make 10 concrete
162. 6:22bets as you saw in the previous slide.
163. 6:24And we're going to see where we end up
164. 6:25at the end of the year. All right. You
165. 6:26know, a lot of new uh engineers that
166. 6:29watch this channel, you guys think I'm a
167. 6:31YouTuber. Uh I'm not. I'm an engineer
168. 6:34and I use this technology every single
169. 6:36day. I bet on it. every single video you
170. 6:38see every you know pattern idea model
171. 6:41experiment like I I'm actively doing
172. 6:45using and looking for the best tools for
173. 6:48the job of engineering right now that
174. 6:50tool is very clearly agents the best way
175. 6:53to get the largest advantage is to show
176. 6:56up before anyone else has even started
177. 6:59those who plan the future tend to create
178. 7:01it sometimes showing up on time is
179. 7:04showing up too late engineering building
180. 7:06businesses Because being inside of
181. 7:08organizations is all about information
182. 7:10advantages. All right? And you want the
183. 7:12largest advantage for the longest period
184. 7:15of time. The best way to do this is to
185. 7:16prepare for opportunity before it
186. 7:19presents itself so that you're ready
187. 7:21when it arrives. Anyway, that's my track
188. 7:23record. I'll leave my previous year's
189. 7:26prediction videos in the description if
190. 7:28you're interested. We're going to cover
191. 7:2925 at the end of this video. Let's jump
192. 7:32in to big bet number one. How are we
193. 7:34going to increase trust in our agentic
194. 7:37systems to become a top 2% engineer?
195. 7:40What would a top 2% engineer do in 2026
196. 7:44that others will not be? Let's start
197. 7:46with a spicy statement here. Anthropic
198. 7:50becomes a monster. I'll tie this back
199. 7:52into trust and agentic systems in a
200. 7:54moment here. But this is the first big
201. 7:56idea I want to talk with you about and
202. 7:58the first big bet I think top engineers
203. 8:00will make. All right, cloud code was
204. 8:02luck. Sonic 3.5 was luck, but everything
205. 8:05since then has been a masterclass in
206. 8:08execution. A lot of engineers don't look
207. 8:10beyond the models, but this company runs
208. 8:12on Nvidia, Google, and AWS. Okay, they
209. 8:16are everywhere. Cloud code unlocked
210. 8:19agentic coding for the masses. This is
211. 8:22the top agentic coding tool and there's
212. 8:24a reason for that. Anthropic isn't
213. 8:26trying to win everyone. They have been
214. 8:28focused on one avatar. That's you and
215. 8:31me. That's the engineer. That's why top
216. 8:332% engineers will continue to bet on
217. 8:36Enthropic. I am betting on this. How
218. 8:38does this tie into trust? It's quite
219. 8:40simple, right? When you bet on one
220. 8:42company with the best track record for
221. 8:45engineers, right? The best tool calling,
222. 8:46the best agent harness, and consistent
223. 8:48execution, your agents will have the
224. 8:51best foundation. And that means you can
225. 8:53just spend less time fighting tooling
226. 8:56and more time shipping. And I just
227. 8:58really want, you know, I put together
228. 8:59this timeline of the insane work that
229. 9:02Anthropic has been doing. Let's just
230. 9:04take a look at some of the top events.
231. 9:05All right, so 2024 June, we had the
232. 9:08Sonic 3.5 release. Huge. One of the best
233. 9:11tool calling models in existence. Okay,
234. 9:14I think the best at the time. Fast
235. 9:15forward February 2025, we got Cloud 3.7
236. 9:19and more importantly, Claude code. This
237. 9:22was the first big introduction for a lot
238. 9:24of engineers. I'd seen it a little bit
239. 9:26before this, but this was the big
240. 9:27introduction to a gentic coding, right?
241. 9:31Context, model, prompt, plus tools,
242. 9:34right? The big three became the core
243. 9:35four. And then fast forward a little bit
244. 9:37more, we got the cloud code SDK. What
245. 9:39does that mean? This is custom agents.
246. 9:41We're going to talk about this more in a
247. 9:42second. Fast forward a little bit
248. 9:43further, we got sub agents. Okay, we got
249. 9:46Opus 4.1, Sonnet 45, and now we have
250. 9:49Opus4. Of course, we have skills. We
251. 9:50have a couple other really important
252. 9:52releases in here as well. But notice
253. 9:54that this is not just about the model
254. 9:56anymore. Okay? Like I mentioned, the
255. 9:59game is not just about the model harness
256. 10:01anymore. It's about what you and I can
257. 10:03do with the technology. It's about
258. 10:05trust. It's about scale. All right? And
259. 10:08a lot of these releases here really
260. 10:10showcase that, right? It's not just
261. 10:13about the model anymore. What's the
262. 10:15model running in? Right? Is it running
263. 10:17in a good harness? Is it caching the
264. 10:19tokens for you? Can you spin up more
265. 10:21compute from that one harness? Right,
266. 10:24sub aents. So, I just want to emphasize
267. 10:26this, right? It's not just about putting
268. 10:28out the best model anymore. It hasn't
269. 10:30been. And you know, Anthropic has been
270. 10:32the leader in this space. Let's make
271. 10:34that absolutely clear. For a Gentics,
272. 10:37for a gentic coding, for a this new
273. 10:39realm that I predicted would emerge
274. 10:41years ago called aentic engineering,
275. 10:44cloud code has been the leader here.
276. 10:46Okay, let's make that super clear. So I
277. 10:48think top 2% engineers are going to
278. 10:50continue to bet on Anthropic as the kind
279. 10:52of leader in the space putting out the
280. 10:55best tooling for this new role of
281. 10:58engineering. All right, let's move on.
282. 10:59Big bet number two in that same vein,
283. 11:01right? Tool calling is the opportunity.
284. 11:05Okay, your agents taking actions on your
285. 11:08behalf is the opportunity. Now this is
286. 11:10something really interesting. Only 15%
287. 11:13of output tokens are tool calls
288. 11:15according to Open Router. All right.
289. 11:16Now, I I want to concretely pull this
290. 11:19up. All right, this is right from their
291. 11:20state of AI. I've read through this and
292. 11:22there's a couple shocking things I want
293. 11:24to mention here. Okay, what one in
294. 11:26particular, check this graph out, right?
295. 11:27This is an obvious great graph.
296. 11:29Reasoning models now represent over half
297. 11:31of all language model usage. And this
298. 11:33makes sense, right? Uh before 01 was
299. 11:36released, one of the first reasoning
300. 11:38models on the channel, we have proof of
301. 11:40this. Uh we were prompt chaining, okay?
302. 11:43We were building our own chains of
303. 11:45reasoning by calling language models
304. 11:48over and over and over making them think
305. 11:50about their answer, right? The the early
306. 11:52versions of what agents are now, right?
307. 11:54Prompts in a loop. Then 01 was released
308. 11:56and we were right. And this turned into
309. 11:58a full paradigm. Okay? Now they take up
310. 12:01over half of all reasoning tokens. This
311. 12:03is fantastic, right? This makes sense.
312. 12:05You scale your compute to scale your
313. 12:07impact. Now, if we scroll down here, we
314. 12:08see something absolutely mind-blowing.
315. 12:10rise of adoption of tool calling. Only
316. 12:1315% of all output tokens were tool
317. 12:16calls. Okay, that's what this is saying.
318. 12:18This is tool invocations over time. Um,
319. 12:20do you see what I see? [laughter]
320. 12:24Do you smell what I smell? This is the
321. 12:27greatest opportunity for engineers
322. 12:29that's going to exist for the next year.
323. 12:31Okay, I don't know about you, but most
324. 12:33of my model usage is tool calling. It
325. 12:37It's raw tool calling. Why is that? It's
326. 12:39because tool calls is a rough proxy for
327. 12:42impact for actions, right? But think
328. 12:43about what a tool call is. It's your
329. 12:45agents taking real actions as you would,
330. 12:48right? Your agents are calling bash
331. 12:51commands. Your agents are reviewing your
332. 12:53work in the browser. They're taking
333. 12:55screenshots. Your agents are now running
334. 12:57arbitrary skills that can call any CLI
335. 13:01script or tool that you build up. Right?
336. 13:03This number is extraordinarily low.
337. 13:06Right? That brings me back to this
338. 13:08opportunity. You can roughly think of
339. 13:10tool calls as impact, right? This is
340. 13:12your agents taking actions. This is what
341. 13:15turns the big three into the core four.
342. 13:18Context, model, prompt, tools. This is
343. 13:20what creates agentic coding. Your agents
344. 13:23are not AI coding anymore. They're
345. 13:25agentic coding. Right? Kind of spoiling
346. 13:28the next bet. Custom agents with custom
347. 13:30tools are the best way to maximize tool
348. 13:33calls and impact. Okay? So I am fully
349. 13:37betting on if you want to be a top 2%
350. 13:39engineer, you will push your agents to
351. 13:42call longer and longer chains of tools,
352. 13:46therefore increasing your total impact.
353. 13:49Tool calling is the opportunity. All
354. 13:52right, that's what this is. This is also
355. 13:54the direct measure of how much you trust
356. 13:57your agents. Okay, the longer the
357. 14:00sequence of correct tool calls is for
358. 14:02whatever problem you're trying to solve,
359. 14:04the more trust you have in your agents.
360. 14:06Okay, I think boil down every bet we're
361. 14:08making here. Boil down everything top 2%
362. 14:10engineers will be doing. It comes down
363. 14:11to this. How long can you let your
364. 14:14agents go? One to many agents, right?
365. 14:16How long can they go before making a
366. 14:19mistake that you have to correct? This
367. 14:21is what I mean when I say trust. Okay.
368. 14:24So then the question becomes, what's the
369. 14:25best way to increase your chance of your
370. 14:28agents calling the right long sequence
371. 14:31of tools? Yes, of course, the model
372. 14:33plays a big role in in that, right? This
373. 14:35is why there are specific models that
374. 14:37consistently hit the top of the
375. 14:39leaderboard. And we can see this, right?
376. 14:40Again, shout out to the open router
377. 14:42state of AI post. Top 10 mostused models
378. 14:45that call the tool. Notice the trend
379. 14:47here, right? It started out with GPT40
380. 14:49mini in the beginning of the year and
381. 14:51then guess which models took over pink
382. 14:54anthropic claw 3.7 and then all of a
383. 14:57sudden Clawude 4 sonnet in blue here
384. 15:00just annihilated this right it took over
385. 15:03the entire market okay and then guess
386. 15:05what happened once again what what what
387. 15:07is this model claude 4.5 and we know
388. 15:10that claude opus is now emerging as the
389. 15:13top best model for long sequences of
390. 15:15tool calls this chart tells is a very
391. 15:18important deep story. Again, as I
392. 15:21mentioned in the first prediction,
393. 15:22Enthropic is creating the best models,
394. 15:25the best tooling for long chains of tool
395. 15:28calls. And that ties right into our
396. 15:30second big bet. Tool calling is the
397. 15:33opportunity. Okay, let's make this super
398. 15:35clear. You know, when everyone talks
399. 15:36about agents, this is what they're
400. 15:38talking about. Reliable long chains of
401. 15:42tool calls. You know, the model is
402. 15:43important, but what comes next? What is
403. 15:45the next most important thing? Again, I
404. 15:47talk about this on the channel all the
405. 15:49time. Everything is the core four. What
406. 15:51puts together the core four? Of course,
407. 15:54your own custom agents. I'm betting big
408. 15:57on custom agents above all. Would you
409. 16:01believe a custom agent with 50 lines of
410. 16:04code, three tools, and a 150 line system
411. 16:06prompt could completely automate a
412. 16:09massive problem you face every day?
413. 16:11Would you believe it if I told you that?
414. 16:13Now, I hope you do because this is true.
415. 16:16I have built these, okay? And I'll be
416. 16:18sharing more of them on the channel, of
417. 16:20course, over 2026. Be sure to like and
418. 16:23comment so the algorithm knows that
419. 16:24you're interested. But this is a real
420. 16:27thing. And this is the value proposition
421. 16:29of custom agents. Why are they so
422. 16:31important? Custom agents solve custom
423. 16:33problems. And custom problems make you a
424. 16:36[ __ ] ton of money. Okay? This is what
425. 16:38this is what engineering is all about,
426. 16:40right? You you you build custom
427. 16:42solutions to solve custom problems that
428. 16:45someone's willing to pay for. Okay. And
429. 16:47custom agents give you that exactly.
430. 16:49They unlock that opportunity with
431. 16:52compute, with prompts, with agents.
432. 16:54Okay. Using agent coding tools is table
433. 16:58stakes now for engineers, right? The
434. 17:00next step up is learning to build custom
435. 17:02agents with some agent SDK. Okay? Every
436. 17:06big AI lab has an agent SDK. Pick one,
437. 17:09lock it in, use it. I use two in
438. 17:12particular. The first one is, of course,
439. 17:14the Clawed Asian SDK. The second we'll
440. 17:16talk about later in the future on the
441. 17:18channel. Top 2% engineers will spend a
442. 17:21ton of time building custom agents. You
443. 17:24know, out of all the things we're going
444. 17:25to talk about here, all the big bets I
445. 17:27believe top 2% engineers are going to
446. 17:29make, I think this is the most important
447. 17:32one. Custom agents above all. All right.
448. 17:35Why? It's because one well-crafted agent
449. 17:38can replace thousands of hours of manual
450. 17:41work. Let me be super clear about this.
451. 17:43Right now, there is a custom agent
452. 17:45running doing someone's job better than
453. 17:49they can. Okay? And it is 50 lines of
454. 17:52code, three tools, and maybe 150 line
455. 17:56system prompt. All right? Now, the trick
456. 17:58is, and we talk about this on the
457. 18:00channel, can you build this? Do you know
458. 18:03what it takes to build this? Can you
459. 18:04prompt engineer? Can you context
460. 18:06engineer? Do you know the right model at
461. 18:08the cheapest price that can let you
462. 18:10build this agent that solves hundreds of
463. 18:13hours of work? All right. When you
464. 18:15fine-tune an agent to solve one problem
465. 18:17extraordinarily well, your trust in that
466. 18:19agent skyrockets. It all comes back to
467. 18:21trust. Okay. The the foundational
468. 18:24problem, right? The limitation is not
469. 18:26the model. It's not the agent, right?
470. 18:28The the capability is all there. It is
471. 18:30our ability to build the right thing to
472. 18:34put together the right context model
473. 18:36prompt tools in our agent to increase
474. 18:39the trust we have in the tools. Okay,
475. 18:42it's all about trust. Now, with that
476. 18:44being said, there are limits to this.
477. 18:47There are limits to what you can do with
478. 18:48a single agent. So, guess what you can
479. 18:50do? You already know what I'm going to
480. 18:52say. The big trend we've been talking
481. 18:54about on the channel, we've explored
482. 18:55this. We have working versions of this
483. 18:58in Agentic Horizon, right? Multi- aent
484. 19:00orchestration. This is the new pattern
485. 19:03that I'm betting big on and I think
486. 19:05other top engineers are looking at this
487. 19:07too as the next frontier in 2026. Multi-
488. 19:11aent orchestration. All right. So in
489. 19:122023 I said one prompt is not enough.
490. 19:15This is before we had any of the clawed
491. 19:18models. This is before we had claude
492. 19:19code. This is before reasoning. Soon
493. 19:21after 01 was released and prompt
494. 19:24chaining aka reasoning became the
495. 19:26standard. So we were right about that.
496. 19:28All right. We saw this trend coming from
497. 19:30miles away.
498. 19:32Now, here's the update. One agent is not
499. 19:34enough. Stop running a single cloud code
500. 19:37instance. Start running 3, 5, 10,
501. 19:40hundreds. Multi- aent orchestration is
502. 19:43the next big trend. This is what's going
503. 19:45to happen next. I'm betting on this. All
504. 19:48right. It's it's for this reason, right?
505. 19:49Tell me who wins. A senior engineer with
506. 19:51one agent or the top 2% running 10
507. 19:54agents. Assuming we're doing legitimate
508. 19:56work with each agent, tell me which
509. 19:58engineer wins. Of course, this is a
510. 20:01stupid question. The engineer with 10
511. 20:03agents wins. Stack that up over a week,
512. 20:05month, and a year. And we're talking
513. 20:07magnitudes of of gapping, [laughter]
514. 20:11you know, of of diffing. We're talking a
515. 20:13difference of magnitudes in the output
516. 20:16from these engineers. Okay? And and this
517. 20:18goes crazy once you start adding time as
518. 20:20as the variable, right? Weeks, months,
519. 20:23quarters, years. Okay, so the trust here
520. 20:26is simple and we've explored these
521. 20:28patterns in tactical agent coding. This
522. 20:30is one of the central thesis of of that
523. 20:32course. More agents gives you more
524. 20:34verification and that gives you more
525. 20:36trust. Right? What one planner or
526. 20:38reviewer will miss three will find.
527. 20:41Okay? Then you can use powerful
528. 20:43techniques like cross validation through
529. 20:45multiple agents to directly increase
530. 20:48your agentic output to directly increase
531. 20:50the trust of your agents. Okay. So
532. 20:54multi-agent orchestration is the next
533. 20:56big trend. For instance, running cloud
534. 20:57code with sub aents. Why was that
535. 20:59release so important? It's because this
536. 21:01was the first look at orchestrating
537. 21:04multiple agents through an agent. Okay?
538. 21:07Because as I've mentioned before on the
539. 21:09channel, you prompt your primary agent
540. 21:11and your primary agent prompts your sub
541. 21:14agents. This is in fact the beginning of
542. 21:17multi- aent orchestration. So we saw
543. 21:19that all the way back here and in
544. 21:21between July and you know the new year
545. 21:23in tactical agent coding and agent
546. 21:25horizon I've shared private versions of
547. 21:27that future right of multi- aent
548. 21:29orchestration. So, this is a big trend.
549. 21:31As you start scaling up your agents, you
550. 21:32quickly run into problems with where do
551. 21:35you put them? Like, what is the proper
552. 21:38space to place your agents? Are are you
553. 21:41really letting them all run rampant on
554. 21:43your computer? A lot of vibe coders are
555. 21:45figuring out that's how you delete your
556. 21:47device, right? And a lot of, you know,
557. 21:49legitimate mid senior plus level
558. 21:51engineers, you know, that there is just
559. 21:53not there's not enough tooling and
560. 21:54safety and and proper scale to place
561. 21:57this many agents right on your device.
562. 22:00And and to be clear, you can do that. I
563. 22:02I do do that. But you have to be very
564. 22:04careful about that. And we can do much
565. 22:06better, right? Where should all these
566. 22:07agents run? They should run in, of
567. 22:09course, agent sandboxes. Okay? Let your
568. 22:13agent run rampid in its own computer.
569. 22:16The top state-of-the-art models by the
570. 22:18big labs can do this. Now, okay, how
571. 22:21does this tie to trust? High trust
572. 22:22agentic engineering sometimes just means
573. 22:25having a space where even if everything
574. 22:27goes wrong, it doesn't matter. Scrap all
575. 22:30of language models. Scrap every agent
576. 22:32you've ever seen. Forget about the
577. 22:33generative AI revolution. Go back to
578. 22:35foundational engineering. This is what
579. 22:36the dev box is, right? This is what the
580. 22:39staging environment is. We're just
581. 22:41preparing and we're building systems of
582. 22:43trust that if something goes wrong here,
583. 22:45it doesn't matter. Only production
584. 22:48matters, right? So, give your agents
585. 22:51their own staging environment, right?
586. 22:53And really, this is a development
587. 22:55environment for each agent. Okay? Your
588. 22:58agents can run in their own computer,
589. 23:00their own sandboxes. And this directly
590. 23:03relates to trust because we just defer
591. 23:05it. I want to solve this bug. Great. I'm
592. 23:08going to spin up 10 agents in their own
593. 23:09computer. Everyone gets a shot at this.
594. 23:11And you don't care about any one of them
595. 23:13except the winner. You defer trust.
596. 23:16Okay. Let the agent prove themselves.
597. 23:18This is the best of end pattern we've
598. 23:20talked about on the channel. This is
599. 23:22what you can use for both green field
600. 23:24and brownfield code bases. All right.
601. 23:26Top 2% engineers will parallelize,
602. 23:29scale, and isolate green field and
603. 23:31brownfield codebases in agent sandboxes.
604. 23:34Let the agent run everything end to end.
605. 23:36This will help you scale your impact and
606. 23:38just straight up defer the trust
607. 23:41entirely until you need it. Right? We're
608. 23:43basically lazy loading the result. Lazy
609. 23:45loading trust in that agent sandbox.
610. 23:48Right? when they've completed the work,
611. 23:50they've exposed the, you know, port or
612. 23:53the result you're looking for for that
613. 23:55training run or for that UI change,
614. 23:57whatever type of work you're doing. You
615. 23:59just pull the result down or you, you
616. 24:01know, this is another uh kind of angle
617. 24:03on this, or you go into their machine
618. 24:05and you look at the result, you see if
619. 24:07you like it, and then finally you put up
620. 24:09that PR or you say, "Go ahead and merge
621. 24:11it." Right? Sandboxes let you defer
622. 24:13trust. Okay? You don't need it until the
623. 24:16very end when you're merging back to
624. 24:18main. We're just scaling up what our
625. 24:21agents can do, right? This is how you
626. 24:22would onboard and and you know train a
627. 24:25new engineer on your team to operate.
628. 24:27You get your own development
629. 24:28environment. Here's how you can run
630. 24:30everything. It's a safe place. We're now
631. 24:32giving our agents that capability
632. 24:34because guess what? No surprise to us on
633. 24:37the channel. Our agents are becoming
634. 24:38more and more capable. All right, big
635. 24:41bet number six. Inloop versus outloop.
636. 24:43What does this mean? Outloop is where
637. 24:45you prompt through some external system,
638. 24:48right? Slack, Discord, GitHub, or your
639. 24:50own personal system. Hint, hint, wink,
640. 24:52wink. And you let your agentic coding
641. 24:54system handle the work. And then they
642. 24:56submit a PR or they deliver some
643. 24:58concrete result for you, right? Maybe
644. 25:00it's a new mock codebase, a prototype, a
645. 25:03UI, or most commonly a PR, right? A pull
646. 25:06request. And then we have inloop agent
647. 25:08decoding. So we have outloop agent
648. 25:10decoding and nloop. So inloop is what
649. 25:12most engineers are doing. Usually
650. 25:13terminalbased agent decoding where you
651. 25:16babysit your agent and worked one prompt
652. 25:18at a time prompting back and forth and
653. 25:20back and forth. Now this is super
654. 25:22powerful but I can guarantee you there
655. 25:24are tasks you are doing that you do not
656. 25:26need to be doing. Top 2% engineers in
657. 25:282026 will have both systems
658. 25:31progressively offloading more tasks into
659. 25:33their outloop agent decoding system.
660. 25:36Okay, this is a massive topic of
661. 25:38conversation inside of tactical agent
662. 25:40decoding. That's my take on how to build
663. 25:42the best systems with agents. In that
664. 25:45course, we hint at a lot of the upcoming
665. 25:47trends in 2026. We get ahead of all of
666. 25:49them. If you want an advantage, check
667. 25:51that out. Link in the description. So,
668. 25:52the idea here is that you'll gradually
669. 25:54build trust in your Outlook system and
670. 25:57you'll hand off more work to your team
671. 25:59of agents that have, you know, more
672. 26:01agents, different task types, more AI
673. 26:04developer workflows, ADWs. And as your
674. 26:07trust grows, you offload more work here.
675. 26:09Okay? This frees up your inloop agent
676. 26:11coding right when you're spending your
677. 26:13precious engineering time and resources
678. 26:15in the terminal or in whatever you know
679. 26:17tool you like to to use. This frees up
680. 26:20your time to really dial into the most
681. 26:22important work. All right. And the goal
682. 26:24here is quite simple. You want to
683. 26:26maximize the number of tasks your
684. 26:28outloop agent coding system is doing and
685. 26:30minimize the end loop. Now many
686. 26:32engineers know by this point there are
687. 26:34tons of cloud tools that can help you do
688. 26:35this. That is a great place to start.
689. 26:37It's not a great place to end. You don't
690. 26:39want an external service operating and
691. 26:41owning the most important resource an
692. 26:43engineer has, their product, their tool,
693. 26:46their core source code, right? You want
694. 26:48to have your own agentic system, your
695. 26:50own agentic ring around your codebase
696. 26:53that operates it for you better than any
697. 26:57cloud tool can. Okay, this is the
698. 26:59central idea inside of tactical agentic
699. 27:01coding. Again, link in the description
700. 27:02if you're interested in checking that
701. 27:03out. Big bet number seven, Agentic
702. 27:06Coding 2.0. Last year I made a bet that
703. 27:09we would see a new paradigm emerge. In
704. 27:122024, AI coding emerged. We were using
705. 27:15tools like Klein, like Ader. We were
706. 27:18just writing code, right? Context model
707. 27:21prompt big three. In 2025, cloud code
708. 27:24brought us a gentic coding agents that
709. 27:26can engineer. They're calling tools,
710. 27:28right? Again, this is the core four
711. 27:30context, model, prompt, tool. Everything
712. 27:33is that. I'm going to say that over and
713. 27:34over again on the channel so it's
714. 27:36crystal clear to you. I don't want you
715. 27:38to get baited [laughter] by any, you
716. 27:40know, feature coming out by any lab or,
717. 27:43you know, tool. If you understand that
718. 27:46fundamental truth that everything is
719. 27:48just the core 4, you will be able to
720. 27:49build and operate at the highest
721. 27:52possible level. All right. Now, I'm
722. 27:54predicting in 2026, top 2% engineers
723. 27:57will push into the next frontier, the
724. 27:59next level of agent coding. I'm calling
725. 28:01it agent coding 2.0. You can call it
726. 28:03whatever you want, but the idea here
727. 28:05ties right into our multi- aent
728. 28:07orchestration. Agents that conduct
729. 28:10agents. The idea is very simple. You are
730. 28:12now talking to your lead agent. Your
731. 28:14lead agent is then spinning up the right
732. 28:16agent for the job, right? Multi- aent
733. 28:19orchestration. This is not
734. 28:20parallelization. Let's be super clear.
735. 28:22That's a part of it. But this is true
736. 28:24orchestration. Okay? You have a lead
737. 28:26agent that can delegate, monitor, and
738. 28:29coordinate. I call them command level
739. 28:31agents, but worker agents is another
740. 28:33term. You can call these whatever you
741. 28:34want, but your lower level agents. If
742. 28:36you're in the cloud code ecosystem right
743. 28:38now, these are just called sub agents.
744. 28:41This is a great place to start, but
745. 28:42there's more work that needs to be done
746. 28:43on the coordination level and the
747. 28:46specialization of your lead agent,
748. 28:48understanding that it can effectively
749. 28:50crud and spawn and delete other agents.
750. 28:53All right, so your orchestrator plans,
751. 28:55spawns, reviews, and ships. Okay. And
752. 28:58this is interesting cuz once again,
753. 29:00we're seeing this trend where we move up
754. 29:01the stack. Again, we're not even the
755. 29:03lead engineer. We're the executive,
756. 29:05right? Because we're just talking to
757. 29:07that lead engineer that runs the
758. 29:10engineering team. Okay? And we're
759. 29:12prompting them and they're taking care
760. 29:13of the rest. Very interesting paradigm.
761. 29:16Okay? So, I'm betting on this. Why is
762. 29:18that? It's for the same reason that
763. 29:19multi-agent orchestration is so
764. 29:20powerful. More agents gives you more
765. 29:22compute, gives you more trust. Your lead
766. 29:24agent understands it, can control other
767. 29:26agents. doesn't solve your problem. You
768. 29:29stop talking to individual agents and
769. 29:31start talking to your lead agent that
770. 29:33conducts the orchestra. You trust the
771. 29:36conductor will deliver the job for you,
772. 29:38right? You trust the system that you've
773. 29:40built. And again, this really hints on
774. 29:42multiple ideas we've been building up
775. 29:44to. Right? Now, we have tool calling on
776. 29:46top of custom agents on top of
777. 29:48multi-agent orchestration on top of our
778. 29:50outloop systems and our inloop systems
779. 29:52when we need to jump in. Right? all
780. 29:54going to turn into this big this big
781. 29:57system, right? Some some like to call
782. 29:59this the swarm. I don't think it's going
783. 30:01to be that big or that, you know, novel.
784. 30:03I think it's just going to be this
785. 30:05pattern. You lead agent command level
786. 30:07agents. Okay? And you might think that
787. 30:10it's just a primary agent and sub
788. 30:12agents. But that's not it. It's agent
789. 30:14orchestration. You need to teach this
790. 30:16agent inside of Agentic Horizon. In that
791. 30:19multi- aent system, the orchestrator
792. 30:21agent aka the lead agent, it itself is
793. 30:24in fact a custom agent, which just hints
794. 30:27at again why custom agents are so
795. 30:29powerful. You can build any type of
796. 30:32agent to solve any problem through tool
797. 30:35calling. All right, let's move on. Big
798. 30:37bet number eight, benchmark breakdown.
799. 30:39So, here is a super super mega benchmark
800. 30:42that proves that the model is the best
801. 30:43at everything. [laughter]
802. 30:46And you can see all the models are at
803. 30:47100%. This is in the future, okay? This
804. 30:50is going to happen. And you know, this
805. 30:52isn't the bet itself, right? The bet
806. 30:54here is that public benchmarks get
807. 30:56saturated. Every model hits 90 to 100%.
808. 30:59They no longer truly give you or I
809. 31:02signal. The models don't prove
810. 31:04themselves on these benchmarks. They
811. 31:06prove themselves to you doing legitimate
812. 31:09engineering work or solving whatever
813. 31:11problem your user has that you solve for
814. 31:14them. All right. So, what are top
815. 31:15engineers doing here? They're going to
816. 31:16build private evaluation systems that
817. 31:19will never become public. Okay? And a
818. 31:22lot of companies, a lot of labs, they
819. 31:24already do this, right? There are
820. 31:26benchmarks they will never share. And
821. 31:28this gives them a massive edge. And so I
822. 31:31think if you're building a company, if
823. 31:32you're building custom agents that
824. 31:34actually matter, that do real userf
825. 31:35facing work, you will measure it
826. 31:38yourself. Not what others tell you, not
827. 31:39what any other benchmark says. you know,
828. 31:41there is very likely a model out there
829. 31:43right now that can do the job of a
830. 31:45custom agent that you're running and you
831. 31:47just have no idea because you do not
832. 31:49have benchmark. So, this is harder.
833. 31:51You're likely paying more for an
834. 31:53expensive model, but also it means that
835. 31:55when a new model comes out that you
836. 31:57should use to get incredible results for
837. 31:59your users, you won't even know it's
838. 32:01there because you have no evaluations.
839. 32:03You have no private benchmarks. The big
840. 32:05bet here is quite simple. It's a
841. 32:07continuation bet that top engineers will
842. 32:10always have private benchmarks to
843. 32:12measure the models for the use case they
844. 32:14care about. Moving on, UIs versus
845. 32:17agents. This is super interesting. Um,
846. 32:19I'm betting that agents are eating SAS.
847. 32:22And I think we're already starting to
848. 32:23see this. Every software as a service
849. 32:25company that doesn't start eating
850. 32:26themselves with agents will be eaten by
851. 32:29a agent focused, agent forward tool.
852. 32:31Okay? And we see this at large with
853. 32:34Google. The Google search bar is a
854. 32:37simple kind of really clear example. Why
855. 32:40would I search and get a list of 10
856. 32:42items when my agent can do this at
857. 32:45absurd scale and speed? So top engineers
858. 32:47will realize agents are the interface.
859. 32:51Many UIs become prompt interfaces that
860. 32:53will look like chatbt. I don't like how
861. 32:56overused the chat interface is. I've
862. 32:58been saying this for over a year. I
863. 33:00would love to see new UI patterns emerge
864. 33:02and I think this bet plays on that.
865. 33:05Okay, a lot of the best UIs are going to
866. 33:07run doubledigit number of agents behind
867. 33:10the scenes to accomplish highly specific
868. 33:12tasks better than any raw deterministic
869. 33:14code could. Top 2% engineers are going
870. 33:17to stop using these slow SAS
871. 33:21applications where you have to click
872. 33:22through these slow UIs to do one thing
873. 33:24that ultimately is just crud on top of a
874. 33:27database. Okay, top engineers are going
875. 33:29to take agents and just build the
876. 33:31minimum tooling. Not even an MCP server,
877. 33:34right? They're going to take tools from
878. 33:35the MCP server, chop them down, and
879. 33:38optimize tool calling so that they get
880. 33:40full control over interacting with that
881. 33:41service at light speeded. And then you
882. 33:43can push that idea into products. Why do
883. 33:46I have to operate this UI? Why can't I
884. 33:48just ask my computer what I want to
885. 33:50accomplish? And that is the UI. Okay.
886. 33:53So, I think we're going to see this
887. 33:54interesting UI versus agents battle
888. 33:57happen in 2026 and top engineers will be
889. 34:00on the side of the agents, right?
890. 34:02Because that gives you once again trust,
891. 34:04speed, iteration, and impact right
892. 34:07inside the UI. Okay, so that's the idea
893. 34:09there. And let's move on to our final
894. 34:11bet here. I believe that the AGI hype
895. 34:13dies. Okay, and we start really locking
896. 34:16into just the thing that matters, which
897. 34:19is agents. Okay, as mentioned in the
898. 34:21beginning, you know, Andrew Karpathy, he
899. 34:23says, "This is the decade of agents." I
900. 34:24completely agree with that. I think
901. 34:26agents are the gift and frankly one of
902. 34:30the final stages of language models.
903. 34:33Now, not to say that there aren't more
904. 34:35compositions to be had, but this whole
905. 34:37AGI ASI stuff, they're not even close.
906. 34:39It's one of the greatest marketing
907. 34:41schemes of all time. I am a massive bear
908. 34:44on this. I I have never cared about this
909. 34:46stuff. It's interesting. It's novel. It
910. 34:48sells very well, right? It it shows up
911. 34:51in headlines really well. But time and
912. 34:53time again, we've seen that this is just
913. 34:55factually untrue and it's been used as a
914. 34:58marketing scheme. Top 2% engineers stop
915. 35:01caring about AGI promises and just lock
916. 35:03into delivering maximum value with
917. 35:05agents. Okay, let's just kill the AGI
918. 35:07hype, right? Let's let's focus on
919. 35:08performing and controlling the things
920. 35:10that we can, which is the best tool of
921. 35:12engineering, agents. Okay, um I'm just
922. 35:15fully sick of this. [laughter] So, and
923. 35:18so I'm betting that other engineers are
924. 35:20really catching on to this and and we'll
925. 35:22just stop caring, right? Like I I don't
926. 35:24respond to, in fact, I'm detracted from,
927. 35:27you know, this AGI ASI hype and talk and
928. 35:31it's a great north star for the deep
929. 35:34researchers inside of these AI labs, but
930. 35:37for everyone else, it's a complete waste
931. 35:39of time. It's vaporware. Okay, let's
932. 35:42focus on shipping useful software. All
933. 35:43right. So, for everyone that made it
934. 35:45here, um, you know, I know this is like
935. 35:47kind of a deep presentational video with
936. 35:49a lot of like conceptual ideas that if
937. 35:51you're not engineering with this stuff,
938. 35:53probably doesn't make any sense, but um,
939. 35:55I have a bonus bet for you here, right?
940. 35:58Number 11. So, this is a hidden bet. If
941. 36:00you just skip through this video, I'm
942. 36:02hoping that you miss this, right? I I I
943. 36:05really want to reward engineers that are
944. 36:07really committed and in planning for
945. 36:09themselves in the future and you want
946. 36:10to, you know, see what other engineers
947. 36:12are planning to improve what they can
948. 36:13do. That's what this is all about,
949. 36:14right? Top 2%, top five, top 10,
950. 36:16whatever you want to, however you want
951. 36:17to frame that, right? I think top 2% is
952. 36:19a great number. That's the elite of the
953. 36:21elite. And then we can go from there,
954. 36:23right? Big bet number 11. We see the
955. 36:25first endtoend agentic engineers emerge.
956. 36:29Now, this is a super edge bet. I'm
957. 36:31probably going to get this wrong, but
958. 36:32I'm betting on this. We're going to see
959. 36:34the first engineer write a blog post
960. 36:36that builds a complete system that
961. 36:39builds the system. So, so this is this
962. 36:40is the this is a central idea we talk
963. 36:43about in tactical agents coding, right?
964. 36:45You you don't want to focus on building
965. 36:47the application anymore. You want to
966. 36:49focus on building the system that builds
967. 36:51the system, right? The system of agents
968. 36:53that does the work for you. So, here's
969. 36:55what we're going to see. the first blog
970. 36:57post where an engineer shares, you know,
971. 37:00some details about their chain of agents
972. 37:03that ship an entire feature end to end.
973. 37:05And I'm talking end to end, prompt to
974. 37:08production, no review, no human in the
975. 37:11loop. This is going to be the pure polar
976. 37:13opposite of those stupid outdated why
977. 37:15can't AI engineer blogs and just like
978. 37:17the the anti- AI crew and engineers who
979. 37:20are just going to get smoked every
980. 37:22single day moving forward from now,
981. 37:24right? And they already have been,
982. 37:25frankly. So end to end agentic engineers
983. 37:27start to emerge. It's going to freak
984. 37:28people out. But you know you and I and
985. 37:31longtime viewers of the channel we've
986. 37:33seen this coming right. In fact once
987. 37:35again it it is inside tactical agent
988. 37:37coding. This is an idea we talk about in
989. 37:39some of the final lessons. This is
990. 37:41coming to engineering. Okay. And in that
991. 37:45framing of trust if you think about it
992. 37:46this is the highest possible level of
993. 37:48trust we can have in our agents. This is
994. 37:51the north star for trusting your agents.
995. 37:53Okay. And in a lot of ways, this is the
996. 37:55endgame of agentic engineering. It's
997. 37:58full autonomy, prompt to production, no
998. 38:00human in the loop. The ultimate
999. 38:02expression of trust in your systems.
1000. 38:04Just like, you know, when your manager
1001. 38:05hands off a feature, a a road map to the
1002. 38:09team and then they just disappear.
1003. 38:12They know the best teams always ship.
1004. 38:15Those are the big bets. Okay. So, this
1005. 38:17is the year of trust. Everything I'm
1006. 38:20doing and everything that I believe top
1007. 38:232% engineers will be doing is centered
1008. 38:27around increasing the trust they have in
1009. 38:30their agentic systems. Right? All the
1010. 38:32way from betting on the right labs,
1011. 38:34increasing their tool calling, building
1012. 38:36out custom agents to solve custom
1013. 38:39problems to make a custom [laughter]
1014. 38:42custom [ __ ] ton of money, right? And
1015. 38:44then you start scaling that up. you
1016. 38:46start building out an orchestration
1017. 38:48agent that can then drive these results
1018. 38:51that you're looking for. All right, once
1019. 38:52you have multi- aent orchestration,
1020. 38:54you're going to want to put all of your
1021. 38:55agents somewhere safe where you don't
1022. 38:58have to care about when things go wrong
1023. 39:00until you want to pick that result,
1024. 39:02right? Best of N. You're not going to
1025. 39:04want to sit in the terminal all the
1026. 39:06time. It's too slow and there are
1027. 39:07problems you just don't need to solve in
1028. 39:10the loop, prompting back and forth,
1029. 39:12babysitting your agent. This is why
1030. 39:13you're going to want to build up an
1031. 39:15outloop system. All right. I think when
1032. 39:17you put together these previous five
1033. 39:19points, you get agentic coding 2.0. You
1034. 39:21get to talk to a lead engineer that
1035. 39:23understands your problem, your codebase,
1036. 39:25everything you're trying to do with a
1037. 39:27bunch of experts aiding it. And then
1038. 39:29that agent spins up the right command
1039. 39:31level agents, right? The right
1040. 39:33specialized agent with their own system
1041. 39:36prompt, right? their own specific
1042. 39:39playbook on how they should best plan,
1043. 39:41build, review, test, and document that
1044. 39:44specific work, right? That all happens
1045. 39:47in a brand new agent coding 2.0 user
1046. 39:51interface, uh, application, product. I'm
1047. 39:54not sure what this looks like, but I'm
1048. 39:55predicting this emerges. We have you
1049. 39:57prompting a lead agent and the lead
1050. 39:59agent commanding one or more agents to
1051. 40:02accomplish the work, specializing every
1052. 40:04agent down to the system prompt, down to
1053. 40:06the user prompt. All right, we we're
1054. 40:08going to have a benchmark breakdown. Top
1055. 40:10engineers in this year of trust, the
1056. 40:12trust disappears [laughter]
1057. 40:15in these public benchmarks. I think
1058. 40:16we're already starting to see this. Top
1059. 40:18engineers just like you and I I have
1060. 40:20this experience, too. I look at some of
1061. 40:21these benchmarks and I go, great. Uh,
1062. 40:23you know, I'm going to block out some
1063. 40:24time. I'm going to run some benchmarks
1064. 40:26I've created and I'm going to really get
1065. 40:27a feel for this for use cases I care
1066. 40:30about. All right, we're going to see UI
1067. 40:31versus agents. Agents are going to eat
1068. 40:33up UIs. You're going to eat up SAS
1069. 40:35applications. If your application is
1070. 40:37just a UI on the database, expect it to
1071. 40:39be cooked. So, start building agents on
1072. 40:41top of it to automatically and
1073. 40:43agentically operate that database for
1074. 40:45you and for your users. And you know,
1075. 40:47number 10, AGI hype dies. We lose trust
1076. 40:50in these labs selling this vision of AGI
1077. 40:53and ASI and we just focus on shipping.
1078. 40:55Okay, so here's what I'm betting my
1079. 40:57time, money, and compute on. I would
1080. 40:59love to know what is your plan for 2026.
1081. 41:01What do you agree with? What do you
1082. 41:03disagree with? Comment down below and
1083. 41:05let me know. Quickly taking a look at
1084. 41:07the 2025 prediction. So, you know,
1085. 41:09here's all the the 15 bets I made. Um,
1086. 41:13AI coding becomes a standard. This is
1087. 41:15true. 84% of developers now use AI
1088. 41:18coding tools. All right, adoption is not
1089. 41:20universal. It's funny to think back at
1090. 41:21the beginning of last year. There were
1091. 41:23still engineers saying things like LLMs
1092. 41:26could never write code. They could never
1093. 41:28do XYZ. Right? Again, this is before
1094. 41:30Cloud Code, Codeex, Gemini, before all
1095. 41:32of that. Right? The first reasoning
1096. 41:34model had just dropped when I was making
1097. 41:36this prediction. Okay? So, we got that
1098. 41:38right. Agentic coding begins. I
1099. 41:40predicted that our tools that were just
1100. 41:43coding would start calling tools
1101. 41:45reliably. Okay? And that's the crux of
1102. 41:48that. If it gets one out of four, right,
1103. 41:51it doesn't matter. That's not good
1104. 41:52enough. But claude code was the primary
1105. 41:54driver. Then we got cursor, agent mode,
1106. 41:56codec, gemini, everyone followed. Right.
1107. 41:58This was true. Nailed that. Right.
1108. 42:00Evergreen AI coding experience. Again,
1109. 42:01we have to give credit to cloud code. It
1110. 42:03redefined agentic coding. Surprisingly,
1111. 42:06it was in the terminal. Um, I was
1112. 42:08expecting this to be a UI, but it was in
1113. 42:10the CLI, and this worked out better than
1114. 42:13anything. So we got a brand new AI
1115. 42:15coding experience in a way. Number two
1116. 42:17and number three are very interlin right
1117. 42:19agent coding did begin and a new coding
1118. 42:22experience to begin inside the terminal.
1119. 42:24All right. Uh the cost of code declines.
1120. 42:26Um not a huge bet to make but this has
1121. 42:29remained true and we're seeing that with
1122. 42:31the new Gemini 3 flash release. The cost
1123. 42:33of code is barreling towards zero. To be
1124. 42:35clear, the right configuration of code
1125. 42:38in a product is still as valuable as it
1126. 42:40was before. In fact, it's more valuable
1127. 42:43because there's a lot more code out
1128. 42:45there that is absolute trash and slump.
1129. 42:48Skill gap earthquake 25% decrease in
1130. 42:50entry-level roles. Senior AI
1131. 42:53orchestrators salaries skyrocketed. I
1132. 42:56got this one right. We see a massive
1133. 42:57decline in entry-le roles. This is still
1134. 42:59continuing throughout the entire year.
1135. 43:01Uh this has been proven true. We get
1136. 43:03points for that looking back. All right.
1137. 43:05LLM model review. So, there is no wall.
1138. 43:08Again, if you remember at the end of
1139. 43:09last year, the end of 2024 when I was
1140. 43:11making this bet, it was very unclear
1141. 43:14whether uh models could push past where
1142. 43:16they were. I bet super hard that they
1143. 43:19could. Again, I'm going to link my
1144. 43:21previous predictions video. It was a
1145. 43:22very different world one year ago in
1146. 43:24December, right? 2024 December was a
1147. 43:26very different world. Everyone was
1148. 43:28talking about the wall, right? I was
1149. 43:30right. There is no wall. Benchmarks have
1150. 43:32been destroyed. New benchmarks have been
1151. 43:34introduced. And guess what's going to
1152. 43:35happen? Those are going to get
1153. 43:36destroyed. I didn't make this a bet for
1154. 43:38this year because I think the writing's
1155. 43:39on the wall. There is no wall.
1156. 43:41Hyperspecialized LLMs got this one right
1157. 43:43as well. We see LLMs, even the best
1158. 43:45ones, start to specialize, right? Cloud
1159. 43:47opus is for engineering. We got this new
1160. 43:50function Gemma that was just released
1161. 43:51specifically for tool calling on
1162. 43:53devices, right? Code stroll. Uh we got
1163. 43:56enterprise models. We got chemistrybased
1164. 43:58models. This is going to continue. Um,
1165. 44:00again, a lot of these bets become
1166. 44:02pointless to continue to make because
1167. 44:04it's just so clear now in the industry
1168. 44:06that there's no alpha here anymore,
1169. 44:09right? There's no alpha in betting in
1170. 44:10this. All right, there's no advantage to
1171. 44:12be had. Okay, year of infinite memory. I
1172. 44:14got this one wrong. I got this one
1173. 44:15wrong. I did not understand this problem
1174. 44:17deeply enough. Context kind of grew. And
1175. 44:20specifically by grow, I mean the
1176. 44:22effective context window of a lot of
1177. 44:23these models did improve, but memory is
1178. 44:26still a huge challenge. I got this
1179. 44:27wrong. All right, we saw GPT4 Ole small
1180. 44:31language models on devices. This is true
1181. 44:3403, Haiku, Quinn. There's a bunch of
1182. 44:37models now that can run on a really
1183. 44:40shitty device all the way down to mobile
1184. 44:43devices. All right, so I got this one
1185. 44:44right. These are probably not the right
1186. 44:46models to list here, but um you can tell
1187. 44:47I had an agent spin that description out
1188. 44:49for me. I predicted uh OpenAI and
1189. 44:51Antropic would release small language
1190. 44:54models. Got this one right. We have
1191. 44:56Haiku and we have uh 03 but also we got
1192. 44:59the GPT OS models right even smaller 120
1193. 45:03billion and 20 billion. All right,
1194. 45:05OpenAI remains number one. I got this
1195. 45:08prediction wrong. They are no longer in
1196. 45:10the top place. Right now it's probably
1197. 45:12between Opus 5 or Gemini 3 in terms of
1198. 45:17the best all-around model for price um
1199. 45:20intelligence and speed. Very clearly the
1200. 45:24winner right now is Gemini 3 Flash as we
1201. 45:26talked about in the beginning. Uh new
1202. 45:28industry-breaking architecture. We now
1203. 45:31have world models and this slowly is
1204. 45:33moving outside of the realm of
1205. 45:35engineering for us. We probably won't be
1206. 45:36covering or talking about these a lot
1207. 45:38unless there's some serious advantage
1208. 45:40for us for modernday uh technology and
1209. 45:44modern day software engineering. But
1210. 45:45certainly we could be. I predicted this.
1211. 45:47It came true. We we now have this new
1212. 45:49world of world models. Exponential slop.
1213. 45:52This came true. This prediction
1214. 45:53describes itself. Uh, big tech shrinks,
1215. 45:56smallmedium businesses grow. I think
1216. 45:58this one's a little more of a mixed bag,
1217. 46:00but ultimately this one's true. You
1218. 46:02could really frame this as correct or
1219. 46:04false if you wanted. I'm going to take
1220. 46:05the credit here. We did concretely see
1221. 46:07big tech cut jobs while small businesses
1222. 46:10were able to thrive with smaller teams.
1223. 46:13Uh, this one's a little weird. If you
1224. 46:14want to dock me on this one, you can. I
1225. 46:16totally understand that. Last last bet
1226. 46:18here that I made last year. Data greater
1227. 46:20than UX, greater than benchmarks.
1228. 46:22Winners have proprietary data and better
1229. 46:24UX and the raw model intelligence was
1230. 46:27commoditized. I think this is right and
1231. 46:29it will continue to be right. The most
1232. 46:30important thing now is your user data
1233. 46:33and the relationship you have with your
1234. 46:35user and access to that data. Then it's
1235. 46:37your user experience. And then lastly,
1236. 46:39it's the language model benchmarks and
1237. 46:41the language model themselves that use
1238. 46:43to feel your your new agentic
1239. 46:44experiences. All right. So, these are
1240. 46:47all the actions I'm taking for 2026. I
1241. 46:51believe this is going to be the year of
1242. 46:52trust. Once again, this is where I'm
1243. 46:54betting my time, money, and compute. Let
1244. 46:56me know what you think in the comments.
1245. 46:57Do you agree? Do you disagree? What am I
1246. 46:59missing here? I'm betting super big on
1247. 47:01these ideas. I have many of these ideas
1248. 47:04already in motion. Again, as mentioned,
1249. 47:06if you're interested, check out Tactical
1250. 47:08Agentic Coding link in the description.
1251. 47:10Many engineers who watch the channel are
1252. 47:12already in there. You're already in
1253. 47:13Agentic Horizon. you're already getting
1254. 47:15the edge on 2026. That's going to be in
1255. 47:18the description if you're interested.
1256. 47:19The big idea here, if I had to pick one,
1257. 47:22it's this custom agents above all when
1258. 47:25it comes to signal, when it comes to
1259. 47:26value, I think this is the one thing you
1260. 47:29can bet big on and you'll see the
1261. 47:32highest return on investment. After
1262. 47:34that, right, once you're building out
1263. 47:36custom agents, you want to start pushing
1264. 47:38into multi- aent orchestration, agent
1265. 47:40sandboxes, and this will naturally kind
1266. 47:42of get you into this next state. You
1267. 47:44want to have that mindset of models
1268. 47:46don't matter anymore. There's no AGI,
1269. 47:48there's no ASI, it's [music] just
1270. 47:51agents. So, let's put my team of agents
1271. 47:53together. Let's build the best custom
1272. 47:55agents. Let's build a great lead
1273. 47:57engineer that helps me communicate with
1274. 47:59my agents. And let's ship. All right?
1275. 48:02Let's get real work done. That's what
1276. 48:03[music] this channel is and will always
1277. 48:05be about. You know where to find me
1278. 48:06every single Monday. Happy holidays.
1279. 48:10Stay focused and keep building.