YouTube2Text

AIに 指示する 時代の終わり|ループエンジニアリング徹底解説 — Transcript

by エンジニア転職チャンネル【RUNTEQ公式】 · 4,159 words · 660 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Mr. Ta, I've been hearing the term "
  2. 0:01harness engineering" a lot lately, and
  3. 0:03our engineers have been working hard on
  4. 0:05putting together harnesses recently,
  5. 0:07right? That's
  6. 0:08right. But I thought it was already
  7. 0:10outdated.
  8. 0:10Huh, harness engineering? We're
  9. 0:12entering the next phase. Actually,
  10. 0:14really? I thought it had just come out
  11. 0:16recently. There's a type of
  12. 0:17engineering that's becoming popular
  13. 0:19right now.
  14. 0:19Loop engineering. And there's another
  15. 0:21new one emerging.
  16. 0:22Yes. It's been a hot topic on X
  17. 0:24recently, and it's also starting to be
  18. 0:26mentioned in places like Anthropik.
  19. 0:28What kind of thing is it? It's
  20. 0:29something called prompter engineering
  21. 0:31itself. You're not doing it. You're not
  22. 0:33even
  23. 0:33giving instructions to the AI?
  24. 0:34Yes. You're not giving instructions.
  25. 0:36How are you doing it? You're having the
  26. 0:38AI give instructions.
  27. 0:40Wait a minute. I can't keep up. The
  28. 0:42concepts are getting more difficult, so
  29. 0:44I, a travel engineer, would like to
  30. 0:46explain it in a way that even beginners
  31. 0:48can understand. Is it
  32. 0:49another development method using AI?
  33. 0:52Yes. I guess it's a method for
  34. 0:54AI-driven development. Group
  35. 0:55engineering. It suddenly started to
  36. 0:58catch on, but this was in June 2026.
  37. 1:00It's been a while since then.
  38. 1:03Everyone started saying exactly the
  39. 1:05same thing. The first person is Peter
  40. 1:07Steinberger. He's the developer of
  41. 1:08OpenClaw, and he's apparently now at
  42. 1:10OpenAir. His idea that we shouldn't
  43. 1:12give prompts to agents anymore, but
  44. 1:14should design loops that give prompts
  45. 1:15to agents, went viral on Flex. The
  46. 1:18second person is Boris Kearney. He's
  47. 1:20the person in charge of the
  48. 1:22Anthropocene Clod Code. Yeah.
  49. 1:24This person said that his job is to
  50. 1:26write loops. He deleted his IDE six
  51. 1:28months ago.
  52. 1:29Yes. What's an IDE? It's a
  53. 1:32development environment. It's an eliter
  54. 1:34or integrated environment for
  55. 1:36development, and you usually enter
  56. 1:38prompts and give instructions there,
  57. 1:40but he's not doing that. The third
  58. 1:43person is D. Oma. He's the director of
  59. 1:46engineering at Google. He named it "
  60. 1:48P-Engineering" on his blog and
  61. 1:50systematized it on his blog. Around
  62. 1:52this time, things started to catch on
  63. 1:54and everyone started talking about loop
  64. 1:56engineering. Yeah.
  65. 1:57People at the forefront of AI are
  66. 1:59saying that if you don't give
  67. 2:00instructions to AI, you shouldn't be
  68. 2:02typing "prompt". That's
  69. 2:03right. It's not about typing prompts,
  70. 2:06it's about designing loops. That's
  71. 2:08where everyone started to move. The era
  72. 2:10of humans giving AI sequential
  73. 2:12instructions is over. From now on, it's
  74. 2:14about designing the mechanism itself
  75. 2:16for giving instructions. I actually did
  76. 2:18group engineering. I tried it myself.
  77. 2:21I'll explain that a little later. It's
  78. 2:22a development style. It's not like this
  79. 2:25concept of loop engineering suddenly
  80. 2:27emerged. It's more like a trend of
  81. 2:30engineering that has been powered up by
  82. 2:32prompt engineering. Prompt engineering
  83. 2:36was the first thing that was said.
  84. 2:38There was also talk about how to use
  85. 2:40blomp skills. Yes. I think this was
  86. 2:42probably the first time it was said,
  87. 2:44and it was said until around 2024, and
  88. 2:46the design concept was to increase the
  89. 2:48quality of each proton. And the next
  90. 2:51one came out in 2025. It's context
  91. 2:53engineering. This contest engineering
  92. 2:55is not just about giving instructions
  93. 2:57as an information environment to show
  94. 2:59to the model, but also about everything
  95. 3:01including the information behind those
  96. 3:03instructions. Harness engineering came
  97. 3:05out in 2026. Isn't it just new? It
  98. 3:09came out this year, but it's the next
  99. 3:11stage. Harness engineering is about
  100. 3:15designing rules, constraints, and
  101. 3:17verifications to see if AI can run
  102. 3:19perfectly. I've been working on harness
  103. 3:22engineering recently, but the new one
  104. 3:24is loop engineering. This means
  105. 3:25creating a mechanism to implement using
  106. 3:28harnesses from June 2026. Yeah.
  107. 3:30Development methods have changed so
  108. 3:32much since the emergence of context
  109. 3:33engineering. I can't keep up anymore. I
  110. 3:35think chat was the basis when we were
  111. 3:37doing front-end engineering and context
  112. 3:39engineering. When we entered context
  113. 3:41engineering, AI agents became the basis
  114. 3:43. And that's where AI agent-based loop
  115. 3:46engineering came in. I don't want you
  116. 3:49to misunderstand that it's not that
  117. 3:51we've switched to a different concept,
  118. 3:53but rather that we've gradually
  119. 3:54incorporated the existing concepts.
  120. 3:56Based on that, it's like wrapping it up
  121. 3:59one more time. Fronts are important,
  122. 4:01contests are important, and harnesses
  123. 4:04are also necessary. Engineering now.
  124. 4:06That's
  125. 4:06right. Originally, context engineering
  126. 4:08was based on prompts, and the question
  127. 4:10was what kind of information to add to
  128. 4:12the prompts. Harness engineering was
  129. 4:14the idea of thinking about how to make
  130. 4:16it run autonomously based on that added
  131. 4:18information. Loop engineering is the
  132. 4:20idea of running it autonomously, not by
  133. 4:23running a single task autonomously, but
  134. 4:25by looping it repeatedly to complete a
  135. 4:27larger task. Isn't that going too far?
  136. 4:31It starts
  137. 4:32with the definition of loop engineering
  138. 4:34. Loop engineering is not about giving
  139. 4:36instructions, but about creating a
  140. 4:38mechanism to give instructions. This is
  141. 4:40the basis. It's the idea of designing a
  142. 4:43loop itself that continues to give
  143. 4:45prompts to the AI coding agent, instead
  144. 4:47of manually typing prompts to it. In
  145. 4:49the past, humans would type prompts.
  146. 4:52The AI would respond, and humans would
  147. 4:54give the next instructions. The AI
  148. 4:55would then work again, and this would
  149. 4:57be repeated. That's the past approach.
  150. 5:00Loop engineering involves humans
  151. 5:01defining the goal. Once the goal is
  152. 5:04defined, A automatically gives
  153. 5:05instructions within that system,
  154. 5:07verifies the results themselves,
  155. 5:09decides on the next task, and continues
  156. 5:11until the goal is achieved. Does this
  157. 5:14mean that
  158. 5:14humans create a system where AI gives
  159. 5:16instructions to AI?
  160. 5:17Yes. I
  161. 5:19think it would be easier to understand
  162. 5:21if we were talking about a team of
  163. 5:22people. Up until now, I was acting as a
  164. 5:25player, like a playing manager. I would
  165. 5:27give instructions to my juniors, verify
  166. 5:29their results, and then say, "Okay,
  167. 5:31let's move on to the next step." The
  168. 5:33entire system for loop engineering
  169. 5:35development is defined here, so please
  170. 5:37create functions within this framework.
  171. 5:40The team thinks and moves forward, and
  172. 5:42that's how it works. Does it
  173. 5:43feel like the management burden has
  174. 5:45increased?
  175. 5:46Yes. It's gotten a lot more involved in
  176. 5:48the management field. It's designing
  177. 5:50how to get the team moving. In cooking,
  178. 5:53there are chefs, stonemasons, and other
  179. 5:55people, but up until now, I was the one
  180. 5:58watching over them and having them help
  181. 6:00me out in the room, but in group
  182. 6:02engineering, the owner would say, "
  183. 6:04Please make this kind of menu," and
  184. 6:06design all the rules for the kitchen.
  185. 6:08The idea is that a chef or something
  186. 6:10like that has become AI and starts
  187. 6:12moving on its own. It's a way of
  188. 6:13thinking about designing a team. When
  189. 6:15people talk about loop engineering,
  190. 6:17they often think of regular guidance
  191. 6:19and retraining. It's easy to mistake it
  192. 6:22for a loop, but loop engineering
  193. 6:23doesn't just have the model repeat the
  194. 6:25same process; it looks at the state and
  195. 6:28decides on the next action. The idea of
  196. 6:31loop engineering is to make decisions
  197. 6:34while looping forward toward the goal,
  198. 6:36rather than doing routine work. Is that
  199. 6:39even possible? It's like giving AI a
  200. 6:42personality, creating a team, and
  201. 6:44creating a state where it can properly
  202. 6:46reduce in a loop. It even sounds scary.
  203. 6:48In fact, it's not just a buzzword that
  204. 6:51came to mind, but has a solid research
  205. 6:53foundation. There are two: the idea of
  206. 6:56react and reflection. React is a
  207. 6:58pattern of thinking, acting, observing,
  208. 7:00and thinking again, repeated over and
  209. 7:02over. In programming terms, it's like
  210. 7:04writing, executing, checking for errors
  211. 7:06, fixing them, and passing the test.
  212. 7:08This idea of repeating this process is
  213. 7:10based on the idea of react. Reflection
  214. 7:12is a pattern of verbalizing and
  215. 7:14recording failures and using them for
  216. 7:16the next try. Based on these, we have
  217. 7:19loop ending. It's a system that builds
  218. 7:21a mechanism for learning from mistakes
  219. 7:23that you can fix on your own. It's like
  220. 7:26systematizing what people do. We've
  221. 7:28talked about the concept, but from here
  222. 7:30on, we'll get a little closer to the
  223. 7:32implementation. There are two layers to
  224. 7:35a loop. I think understanding this will
  225. 7:38help you understand a lot better. The
  226. 7:41two are the inner loop and the outer
  227. 7:42loop. It's a system where you verify
  228. 7:46things yourself within a single task.
  229. 7:48Is
  230. 7:49this the inner loop?
  231. 7:50Yes. I think many of you often have
  232. 7:54complaints about AI, like when you ask
  233. 7:56it to fix something, the AI says it's
  234. 7:58done. But when you look at it, it's not
  235. 8:02fixed or it's still broken. You can
  236. 8:04verify that by running a loop. You
  237. 8:07don't just say it did it, you have it
  238. 8:08check to see if it's fixed?
  239. 8:10Yes. You edit, create a test, run the
  240. 8:12test, detect failures, detect other
  241. 8:15patterns, fix them, re-run it, and
  242. 8:17check if it passes. This loop is
  243. 8:19repeated. The
  244. 8:21inner loop is the one that verifies and
  245. 8:23fixes things until it's complete. In
  246. 8:25the conventional format, people do it,
  247. 8:28people give instructions, AI does it,
  248. 8:30and people check it. Instead, people
  249. 8:32give instructions, AI executes it, AI
  250. 8:34checks it, and if it doesn't work, it
  251. 8:36goes back to the previous AI, and
  252. 8:38that's repeated all the way to the pole
  253. 8:40and finally goes back. This is what's
  254. 8:42called the inner loop. The key point
  255. 8:45here is to do verification in detail.
  256. 8:48For example, it's better to clarify
  257. 8:50what you want to verify. Even if you
  258. 8:55create something that only performs
  259. 8:57testing, there are other things like
  260. 8:59security risks, code errors, and slight
  261. 9:01inconsistencies in the specifications,
  262. 9:03so if you review each of them, the
  263. 9:05verification will be more effective.
  264. 9:09Are there many AIs that perform testing
  265. 9:11and AIs that check security? Yes. By
  266. 9:14doing that, it's like you can't proceed
  267. 9:16to the next gate after four gates. Yes.
  268. 9:19It's an AI team. Yes. Next is the outer
  269. 9:22loop. The current inner loop is a small
  270. 9:24single task. Since this single task
  271. 9:26alone cannot meet the requirements, you
  272. 9:28line up the functions and run them all
  273. 9:31over and over again. It's like running
  274. 9:33a session. If it takes one session to
  275. 9:36create one function, and you continue
  276. 9:38to do that over time, and gradually
  277. 9:40move on to the next task, you're
  278. 9:42creating a loop, right? Yes, that's
  279. 9:45right. The point is to learn more by
  280. 9:47repeating. If there was a mistake in
  281. 9:51the first task, you can design it so
  282. 9:53that you can apply the lesson to the
  283. 9:55next task,
  284. 9:56and then you'll be able to use that
  285. 9:58lesson to talk about different tasks.
  286. 10:00It's like properly
  287. 10:02passing on the PDC. It's like designing
  288. 10:04to avoid repeating the same problem.
  289. 10:06Yes. It's like developing people. In
  290. 10:09Loop Engineering, written by Ali
  291. 10:11Diomani, there's a concept called the 5
  292. 10:13+ 1 components that make up a loop. He
  293. 10:16organized it so that a properly
  294. 10:17functioning loop consists of five
  295. 10:19components plus memory. I think you
  296. 10:21might not understand what I'm talking
  297. 10:23about, so let me explain. The first is
  298. 10:26automation. The role of automation is
  299. 10:29to periodically discover tasks and
  300. 10:31start agents. This is the core of the
  301. 10:34loop. The second is the work tree. It's
  302. 10:37the idea of properly separating work
  303. 10:40areas so that multiple agents don't
  304. 10:42collide. So that two AIs don't do the
  305. 10:44same thing, right? Well, if they work
  306. 10:46in the same place, the environments
  307. 10:47will clash, so let's create an
  308. 10:49environment where they can work in
  309. 10:50parallel. The third is skills. This
  310. 10:54project will be permanently persisted,
  311. 10:56which might be the idea behind
  312. 10:58harnesses. The fourth is flag
  313. 11:00connectors. Things like Habooslack DV.
  314. 11:04Let's prepare things to connect to
  315. 11:06dependent tools and properly connect to
  316. 11:08the outside world. The first is
  317. 11:11subagents. As I mentioned earlier,
  318. 11:14these are necessary for the division of
  319. 11:16labor in a team, so let's separate the
  320. 11:18roles of creation and verification. And
  321. 11:20then there's the plus 1, memory. The
  322. 11:23key point is that the state is
  323. 11:24persisted outside the conversation, not
  324. 11:26within the session. If we process
  325. 11:31something within the session, the
  326. 11:33processing ends there, so the idea is
  327. 11:35to have it persisted externally. For
  328. 11:38example, let's record any missing
  329. 11:40failures outside the task, not within
  330. 11:42it. That's
  331. 11:43right. This is the 5 + 1 approach.
  332. 11:45Among these, I think the fifth point,
  333. 11:47the subagent, is something that needs
  334. 11:49to be designed carefully. It's
  335. 11:52closer to the idea of a harness than a
  336. 11:54subagent loop, but in order to ensure
  337. 11:56the quality of what AI creates, it's
  338. 11:58very important to separate the roles of
  339. 12:00the agents. I think I talked about the
  340. 12:03harness in the anthropics before, but
  341. 12:05it's important to clearly separate the
  342. 12:07roles of creator and checker. In terms
  343. 12:09of what I've talked about on this
  344. 12:11channel, first there's the planner. The
  345. 12:13planner makes the plan, then there's
  346. 12:15the generator. The implementer writes
  347. 12:17the code, and the elevator, the checker
  348. 12:19, checks and provides feedback, and
  349. 12:21this cycle goes on and on. If you work
  350. 12:23on this part carefully, the quality
  351. 12:25tends to improve. In a previous paper,
  352. 12:27there were results showing that
  353. 12:29dividing the roles rather than having a
  354. 12:30single agent do it leads to better
  355. 12:32results. That's
  356. 12:33right. Designing the roles of the team
  357. 12:35is key. Another thing is to have a
  358. 12:37state, but since it's passed on to the
  359. 12:40subagents one after another, even if a
  360. 12:42subagent makes a mistake, it won't be
  361. 12:44passed on to the next one. That means
  362. 12:46that by writing out the failed tasks in
  363. 12:48the subagent to an external file, you
  364. 12:50can write out the status to the file to
  365. 12:52indicate whether it failed or succeeded
  366. 12:54and move on to the next task. If you
  367. 12:56make sure to write out completed tasks,
  368. 12:58tasks in progress, and things learned,
  369. 13:00the quality will gradually improve. So
  370. 13:03that's how far you've
  371. 13:04come in your work. It's like you're
  372. 13:06designing an organization that lets you
  373. 13:08learn and not let the next one go.
  374. 13:09That's
  375. 13:10right. I'll
  376. 13:11tell you something even more important.
  377. 13:14Some people might think that the loop
  378. 13:17is a place where some amazing genius,
  379. 13:19like AI, acts on its own, but to put it
  380. 13:21simply, the loop, the team that I
  381. 13:23mentioned earlier, is not really run by
  382. 13:26super engineers. It's more like a group
  383. 13:29of juniors who work tenaciously. I see.
  384. 13:31If the instructions are vague, they'll
  385. 13:33get together, and if the goal is not
  386. 13:35clear, they might go in the wrong
  387. 13:36direction, and if you don't set any
  388. 13:37restrictions, they'll just keep trying
  389. 13:39endlessly. So it's like having a lot of
  390. 13:42juniors. That's right. It's good to
  391. 13:44think of it as a team of juniors. It's
  392. 13:46like defining the rules clearly. There
  393. 13:48are four important things to keep in
  394. 13:49mind when defining rules. The first is
  395. 13:51a clear task definition. Can you
  396. 13:53clearly state when it's complete in one
  397. 13:55minute? Make sure that everyone
  398. 13:57understands what your goal is. If the
  399. 13:59goal definition is vague, the outcome
  400. 14:01will also be vague. The second is a
  401. 14:03safe work environment. As I mentioned
  402. 14:05in Work 3 earlier, the design should be
  403. 14:07such that even if something goes wrong,
  404. 14:09the environment is isolated and the
  405. 14:11whole thing won't break. It's better to
  406. 14:13design this in advance. The third is
  407. 14:16whether or not you can see the success
  408. 14:17visually. This means using tests,
  409. 14:19linting, and type checking to
  410. 14:21automatically determine the overall
  411. 14:23cost. It's like having a solid cariness
  412. 14:25system in place. By doing this, it
  413. 14:26becomes easier to see whether or not
  414. 14:28something has clearly been successful.
  415. 14:30The fourth is a hard stop. This
  416. 14:32prevents infinite loops and runaways.
  417. 14:34If you keep working until this quality
  418. 14:36is met, you'll just keep working
  419. 14:38endlessly. You won't have any idea what
  420. 14:40the outcome will be. You need to set it
  421. 14:42up so that if you get stuck, you can
  422. 14:44ask your superiors. It's something like
  423. 14:47including the maximum number of retries
  424. 14:49, how much time it should take, and if
  425. 14:51the cost exceeds a certain level, stop,
  426. 14:54or confirm. To prevent them from
  427. 14:56running too fast and collapsing,
  428. 14:58juniors tend to take on too much on
  429. 15:00their own and keep doing tasks
  430. 15:02indefinitely, right? I think there are
  431. 15:05rules like "if you're having trouble
  432. 15:07with time, ask someone" and so on. In
  433. 15:10the same way, it's important to clearly
  434. 15:12define what and how much you should get
  435. 15:14stuck on and what you should fail at in
  436. 15:16the conjunctive case. I thought it was
  437. 15:18about team and organizational
  438. 15:20management. Since we were talking about
  439. 15:23next year, I'll actually talk about the
  440. 15:25loop I created from here. In general
  441. 15:27terms, I created a skill called Epic
  442. 15:29Flow for the outer loop and a skill
  443. 15:31called Transfer Flow for the inner loop
  444. 15:33. I'll start by explaining the inner
  445. 15:35loop transfer flow. Since it transfers,
  446. 15:37the implementer implements it. There is
  447. 15:39also a tester as another subagent. The
  448. 15:41tester tests. If it passes, the next
  449. 15:43step is a multi-axis review. The
  450. 15:47reviewers I created are a usage checker
  451. 15:49, a code checker, and a security
  452. 15:51checker. There are three reviewers.
  453. 15:54Once it passes the four-axis review, it
  454. 15:56is returned to the implementer, who
  455. 15:59then submits a full request. Let me
  456. 16:01explain in more detail. The one-cycle
  457. 16:04flow is called the epic flow from the
  458. 16:06middle loop. An epic is a large feature
  459. 16:09that can be given a name. When you
  460. 16:12start the pick flow and enter the
  461. 16:14requirements there, the planner is
  462. 16:15started first. The planner breaks down
  463. 16:18the major requirements into
  464. 16:19sub-requirements. In creating the
  465. 16:21sub-requirements, dependencies,
  466. 16:23acceptance conditions, and objectives
  467. 16:25for each sub-requirement are determined
  468. 16:27. There are several things to do to
  469. 16:29create a large feature, so the runner
  470. 16:31breaks them down and determines them.
  471. 16:33Is this okay for now? Here, we will
  472. 16:35send in a merchant. When the human
  473. 16:37merchant
  474. 16:38sends the merchant, the main agent will
  475. 16:40give instructions one by one for each
  476. 16:42cycle. The one-cycle flow we saw
  477. 16:43earlier is used when giving
  478. 16:45instructions. In the one-cycle flow,
  479. 16:46one task is completed one after another
  480. 16:48.
  481. 16:48Yes. The one-cycle flow is created, and
  482. 16:50when it comes to completing tasks, the
  483. 16:52main agent gives instructions to
  484. 16:53complete them in the one-cycle flow.
  485. 16:55The
  486. 16:55main agent passes one task to the
  487. 16:57developer and implementer. After one
  488. 17:00week, we come back. Then, we make it an
  489. 17:03epic branch. By repeating this process,
  490. 17:07this entire function is completed.
  491. 17:10Finally, this entire epic pull request
  492. 17:12is completed, so a human checks it one
  493. 17:14more time, and then the initial budget
  494. 17:16and completion checks are done. This is
  495. 17:19the only part that humans do. The AI
  496. 17:22checks whether each of the smaller
  497. 17:23decomposed parts is done. If it's OK,
  498. 17:27it moves on to the next loop. The
  499. 17:29finished product is then checked and
  500. 17:31approved. That's right. The check is
  501. 17:34done on the assumption that the quality
  502. 17:35of the work is guaranteed. Is there a
  503. 17:37chance that the finished
  504. 17:38product will be far from what was
  505. 17:40expected? The
  506. 17:41accuracy is incredibly high.
  507. 17:42Yes, that's right.
  508. 17:43Yes. First, we hand over the major
  509. 17:45requirements, but if there are any
  510. 17:47ambiguous points, they ask questions.
  511. 17:49The planner asks questions, and this is
  512. 17:51what happens. Because we refine the
  513. 17:53requirements. Once they understand, the
  514. 17:55planner starts working on the task over
  515. 17:57. Because that's done well, a proper
  516. 18:01product is created.
  517. 18:02Yes. Yes, that's right. Also, under
  518. 18:04this main part, we have a harness
  519. 18:06properly drawn, so the quality check
  520. 18:08method and design are all handed over
  521. 18:11in advance in the harness, so it's even
  522. 18:13harder to get lost.
  523. 18:14How much time does it take? In the
  524. 18:16current phase 123, the plan can be
  525. 18:18completed in 5 to 10 minutes, but from
  526. 18:20there, the development confirmation
  527. 18:22takes quite a while, depending on the
  528. 18:24size of the previous day.
  529. 18:25Yeah. Yeah. Yeah.
  530. 18:27And we often hit the limit. As
  531. 18:28Hikomo-san said, "I hit the limit again
  532. 18:31." Recently, I've been making an epic
  533. 18:33before going to bed, executing it, and
  534. 18:35waking up in the morning, and it's
  535. 18:37finished. A great fairy gave me a shot.
  536. 18:39As I mentioned earlier about the
  537. 18:41planner, there are 6 sub-agents. They
  538. 18:43are planning, implementation testing,
  539. 18:45usage confirmation, and security
  540. 18:46confirmation. I'll tell you what kind
  541. 18:48of design each one has. Earlier, we
  542. 18:50were talking about using Q a lot, right
  543. 18:52? We use different models. The planner
  544. 18:55uses Opus. This involves breaking down
  545. 18:58tasks from the epic, determining
  546. 18:59dependencies, acceptance criteria, and
  547. 19:01areas of concern, as well as preparing
  548. 19:03objectives and security checks. There
  549. 19:05are some areas where security is not
  550. 19:07necessary, so we design here to
  551. 19:08determine whether it is necessary or
  552. 19:10not. Next is the implementation team.
  553. 19:12We use Sonnet for this. The
  554. 19:14implementation team has put together a
  555. 19:16solid harness, and we ask them to
  556. 19:17proceed with the description of the DD
  557. 19:19bull request based on the harness. The
  558. 19:21tester. This is also one of them. They
  559. 19:23check the implementation based on the
  560. 19:25implementation criteria. This is not
  561. 19:27just a single debug, but we also have
  562. 19:29them check the screen using a
  563. 19:30playwright. They then check whether it
  564. 19:33is okay, so to check whether the
  565. 19:35operation is working properly, it is
  566. 19:37not just the code but also the
  567. 19:38operation. We use a playwright to
  568. 19:42launch the screen in a browser and
  569. 19:44check it. The usage checker. We set the
  570. 19:48implementation criteria, but they check
  571. 19:50whether the implementation criteria are
  572. 19:52sufficient for the purpose and intent,
  573. 19:54and whether the scope of impact has
  574. 19:56spread to other areas. If there is a
  575. 19:58gap, we say that it is missing and
  576. 20:00return it. The last is the code
  577. 20:02reviewer. This is also done using
  578. 20:04Sonnet. If there are no issues, we'll
  579. 20:06just leave it as is. If there are any
  580. 20:08issues, we'll point them out right away
  581. 20:10. It's the same as human code reviews.
  582. 20:15Security checks. This is important, so
  583. 20:18we've used Opus. We're keeping a close
  584. 20:20eye on overall security, but we try not
  585. 20:22to start it up where it's not needed,
  586. 20:24and it only starts up when we're
  587. 20:26writing code. So, each of the six
  588. 20:29subagents starts up. This is how we set
  589. 20:32up an independent response. If I start
  590. 20:35talking about the details, well, there
  591. 20:37are things I want to say, but I
  592. 20:38gradually lose track of what I'm saying
  593. 20:40, and some people might think, "We
  594. 20:42don't need engineers anymore." But
  595. 20:44actually, when I ask juniors, they ask
  596. 20:46me how I can put together so many
  597. 20:48people. As I've been saying, it's
  598. 20:49because I have experience running teams
  599. 20:51of people. I've been working on a lot
  600. 20:54of development-style flows, such as how
  601. 20:56to make teams work when working on
  602. 20:58projects. So it just came naturally to
  603. 21:01me.
  604. 21:01Yeah. That's what AI has replaced. Well
  605. 21:04, when you think about the skills
  606. 21:05required for a loop engineer to work
  607. 21:07together, the first thing is how to
  608. 21:09create software as a GM. There are also
  609. 21:11a lot of things to consider, so if you
  610. 21:12learn these things, it will be easier
  611. 21:14to understand. Then there are project
  612. 21:17management and testing techniques. You
  613. 21:21also need to think about how to ensure
  614. 21:23quality. If you have upstream skills,
  615. 21:25it will be easier to work together. On
  616. 21:28the other hand, you are someone who has
  617. 21:30been running a team in software
  618. 21:31development. I think it would be
  619. 21:34difficult to know what kind of roles to
  620. 21:36create and what kind of loop to use. I
  621. 21:39think there are many things that need
  622. 21:41to be added, such as whether to perform
  623. 21:43verification or define a design
  624. 21:45document. However, it is not enough to
  625. 21:47think that this is the only correct
  626. 21:48loop. In software development, the
  627. 21:51method changes slightly depending on
  628. 21:53the site and requirements, and it needs
  629. 21:55to be flexible, so the way to work
  630. 21:57together needs to be considered
  631. 21:59according to each project. If you are
  632. 22:02going to work together all the time,
  633. 22:04you need to create a situation where
  634. 22:06people can run it. That is why studying
  635. 22:08is important.
  636. 22:09If anyone watching this is interested,
  637. 22:11can I consult with you?
  638. 22:14In the future, I'm thinking of reviving
  639. 22:16live streaming and such, so please feel
  640. 22:18free to ask me there as well. Also, for
  641. 22:21those who watched this video, I've
  642. 22:23created a small e-book called Loop
  643. 22:25Engineering. Yes. That's right. I'll
  644. 22:29hand it out, so if you register on LINE
  645. 22:31, I'll send it to you as soon as
  646. 22:32possible, so if you want to know more,
  647. 22:34please register on LINE. Yes. So you've
  648. 22:36even created an e-book? Yes. I've
  649. 22:39recently become a bit addicted to
  650. 22:40creating e-books. It's like a
  651. 22:42new hobby, so if you're interested,
  652. 22:44please register on LINE and check out
  653. 22:46the book. Yes. Mr. Hon, am I cut out to
  654. 22:49be an engineer too?
  655. 22:49If you're not sure, I'll take the test.
  656. 22:51I'd like to
  657. 22:52try it. You can easily take the test
  658. 22:53from LINE. Everyone, please try the
  659. 22:55engineer aptitude test from LINE in the
  660. 22:57description.

About this transcript

This page contains the full transcript of AIに 指示する 時代の終わり|ループエンジニアリング徹底解説 by エンジニア転職チャンネル【RUNTEQ公式】, generated from the public captions YouTube serves with the video. The transcript has 4,159 words across 660 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.