Deepseek lo hizo: ¡liberó SuperAgent GRATIS! ¿el superagente para 2027? — Transcript

Full transcript

1. 0:00DeepSeek, the most powerful Chinese lab
2. 0:03in the world, just launched Super Agent
3. 0:06, an agent that can monitor your GitHub
4. 0:09and all your team's chats, working on
5. 0:11several tasks at once without you
6. 0:14needing to babysit it.
7. 0:16It is a prototype, and right now it has
8. 0:19just gathered over 18,000 stars in about
9. 0:224 days on GitHub. I wanted to see if
10. 0:26this super agent was actually real, so I
11. 0:29repeated some of the main experiments
12. 0:32from the prototype in a small use case
13. 0:35with a model other than DeepSeek. The
14. 0:38result, honestly, surprised me, because
15. 0:40here comes what I think no one is saying
16. 0:43about the DeepSeek super agent.
17. 0:46A single sentence in your prompt can
18. 0:48provide the same result as their entire
19. 0:51architecture, saving you perhaps
20. 0:54millions of tokens.
21. 0:56And if this is your first time on my
22. 0:57channel. I am Vera, and today I'll tell
23. 0:59you what that sentence is, what ideas
24. 1:01from DeepSeek you can copy today without
25. 1:03installing absolutely anything, and why
26. 1:05I still wouldn't install their prototype
27. 1:07on your computer.
28. 1:09The main idea behind this prototype is
29. 1:11that the limit really isn't the agent,
30. 1:14it isn't the model; the limit is you.
31. 1:17And for this, I have to explain a little
32. 1:19bit about how I work myself.
33. 1:21I have one coding session fixing, for
34. 1:23example, a bug, another writing maybe
35. 1:25some design or working on the UI of a
36. 1:27landing page, something like that.
37. 1:30Another session, maybe working on a pull
38. 1:32request, and another that I don't even
39. 1:34remember why I opened. And constantly I
40. 1:37ask myself, which one did I tell what?
41. 1:39DeepSeek put a number on that.
42. 1:42A person can follow at most three to
43. 1:45five sessions with an agent at a time.
44. 1:48Much more than that. Really makes you
45. 1:50lose the thread of the conversation.
46. 1:52Most people cannot handle more than that
47. 1:55. And their conclusion is the idea for
48. 1:58this whole video. A single task is
49. 2:00limited by the model, but several tasks
50. 2:02at once are limited by our attention.
51. 2:05What holds you back is no longer the
52. 2:07model, it is really you. Jumping from
53. 2:10tab to tab. So before we continue,
54. 2:12answer me in the comments how many agent
55. 2:14sessions you have open right now, just
56. 2:16the number. I have up to six right now.
57. 2:20What does this mean? What DeepSeek
58. 2:21proposes is best understood with an
59. 2:23example. So let's follow a bug from
60. 2:25beginning to end. Someone on your team
61. 2:27writes in the group, "sometimes the
62. 2:29screen goes blank."
63. 2:30The first thing that happens with this
64. 2:32prototype is that the agent is not going
65. 2:34to read that message using an expensive
66. 2:36model. Before that, it passes through a
67. 2:38code filter: who wrote it, in what group
68. 2:40, and if it's their turn to respond.
69. 2:42Most of the daily messages stay there.
70. 2:46Only what passes the filter reaches the
71. 2:48model, and that's where the concept of
72. 2:51the router comes in.
73. 2:53Think of it like a workshop reception.
74. 2:56It listens, decides who’s in charge, and
75. 2:59never picks up a tool.
76. 3:01It hands the task to a worker, who is
77. 3:03like a mechanic with their own bench, a
78. 3:05single task, their own context, and all
79. 3:08the tools.
80. 3:10But here comes the trick.
81. 3:13The worker opens a pull request with the
82. 3:15fix, and someone comments on the PR: "
83. 3:18we need a test."
84. 3:20That comment doesn't go back to the
85. 3:21reception. In reality, it goes to the
86. 3:23director of that worker for that same
87. 3:25task because they subscribed to that
88. 3:28thread.
89. 3:29This way, the router never gets
90. 3:30overwhelmed. And you never talk to five
91. 3:33agents at once. The workers report to
92. 3:35the router, and the router is the only
93. 3:38one that answers you. To you, it’s
94. 3:40always the same agent in the chat or on
95. 3:43GitHub. And the most proactive thing is
96. 3:46that DeepSeek uses a "reproduce and fix
97. 3:48in silence" rule.
98. 3:50If someone mentions a bug, the agent
99. 3:52tries to reproduce it on its own and
100. 3:54only speaks if it confirmed it and has a
101. 3:57fix. If not, it doesn't answer you. But
102. 4:00all this sounds a bit theoretical.
103. 4:03What things can you do right now using
104. 4:06any harness from this DeepSeek prototype
105. 4:08?
106. 4:09What you can do with this super-agent
107. 4:11philosophy would be the following. Mind
108. 4:14you, I have to clarify something
109. 4:16important: none of these ideas or the
110. 4:19functioning of the prototype itself
111. 4:21depend on using DeepSeek models.
112. 4:25You can use this today with your
113. 4:27sessions in Claude, Codeium, Codex,
114. 4:29Hermes, or any other agent you're using
115. 4:32.
116. 4:34The first thing you can do is, one, your
117. 4:36main session shouldn't execute; use it
118. 4:39only to talk to you and distribute work
119. 4:42as a router.
120. 4:44The long stuff goes to sub-agents or
121. 4:46other sessions. Even their prompts are
122. 4:49separated like that. And for these
123. 4:51different sessions or sub-agents, you
124. 4:54have to distinguish or differentiate
125. 4:56even the prompts. One, for example, to
126. 4:59admit messages, another to distribute
127. 5:01them, another for the worker that will
128. 5:04work on each of the problems, issues, or
129. 5:06bugs.
130. 5:07Two, one task, a single context. Don't
131. 5:10ask the same session to fix the bug.
132. 5:12Write the landing page and review the PR
133. 5:14; keep each thing in its own session and
134. 5:17have it notify you when it finishes each
135. 5:19task.
136. 5:20Three, filter before thinking.
137. 5:23If your agent receives notifications or
138. 5:26messages, set up simple rules before the
139. 5:29model using hooks or a single script
140. 5:31that ignores them and only responds if
141. 5:34you are mentioned for approval.
142. 5:37This saves millions of tokens and a lot
143. 5:40of noise.
144. 5:41Four, have people talk to you only when
145. 5:44necessary and briefly.
146. 5:47In their chats, DeepSeek limits each
147. 5:50response to 100 words, which is more
148. 5:52than enough to understand each of the
149. 5:55steps it is executing from the plan or
150. 5:57assigned task.
151. 5:59But all this sounds quite good on paper
152. 6:01, but does it really work?
153. 6:04Well, DeepSeek put it to the test,
154. 6:06simulating a team chat with about 300
155. 6:08messages, work requests, chats,
156. 6:10follow-ups, and cancellations.
157. 6:13They compared their superagent against a
158. 6:15single session capable of creating
159. 6:17subagents. The programming work involved
160. 6:19waiting, so they were measuring
161. 6:21coordination, not code.
162. 6:23For that single session, they tested two
163. 6:25prompts: a neutral one, for example,
164. 6:27with a message like "use a subagent when
165. 6:30appropriate," and another proactive one
166. 6:32with a message like "prefer giving each
167. 6:34task to a subagent."
168. 6:37With the neutral prompt, the model never
169. 6:39delegated; it did everything itself, new
170. 6:41messages interrupted it, and its slowest
171. 6:43tasks took nearly twice as long.
172. 6:46In contrast, using the superagent's
173. 6:48proactive prompt, delegation improved,
174. 6:50but not entirely.
175. 6:52In one out of five runs, the model
176. 6:54decided the work was, quote-unquote,
177. 6:56small, and did it alone. That is one
178. 6:59thesis. With a single session, you
179. 7:01depend on the model listening to you;
180. 7:03with their architecture, you don't. That
181. 7:05said, they admit it themselves. The
182. 7:07subagent consumes many more tokens. You
183. 7:10trade tokens for time, and that is
184. 7:12precisely the trade-off.
185. 7:14And when we see how this whole flow
186. 7:16works, we have to ask ourselves
187. 7:17something equally uncomfortable: where
188. 7:19do we fit in?
189. 7:21According to DeepSeek, if the system
190. 7:23handles sessions and context, you move
191. 7:25up, but you stop executing and start
192. 7:27deciding and approving.
193. 7:28Their repo is based on their philosophy
194. 7:31. Take as little human attention as
195. 7:33possible and provide the greatest value
196. 7:36possible.
197. 7:37Every message, every notification, and
198. 7:39every question from people must be worth
199. 7:40more than the attention it costs you.
200. 7:42If not, it gets removed. That is why I
201. 7:45say the limit is really you.
202. 7:48Not because you get in the way, but
203. 7:49because your attention is the most
204. 7:51expensive part of this entire system.
205. 7:54And now for the part no one is telling
206. 7:56you: DeepSeek's Superagent is a research
207. 7:59prototype, and they acknowledge and
208. 8:01emphasize this themselves; it is not one
209. 8:04of their products.
210. 8:06They state it right at the top of the
211. 8:09entire README. And this is truly serious
212. 8:11.
213. 8:12In their installation guide, they state
214. 8:14that workers run with full access as
215. 8:16your system user and can execute
216. 8:18programs and read your personal folder.
217. 8:21If you test it, do so with a separate
218. 8:23account, using only limited credentials
219. 8:26.
220. 8:27Also, you must keep in mind that a
221. 8:29different agent that I don't use, but
222. 8:31that can read chats and comments from
223. 8:34anyone, is exposed to prompt injection.
224. 8:39Someone hides instructions in a message
225. 8:41, the agent follows them with access to
226. 8:43your machines, and this means an
227. 8:45attacker can easily access your
228. 8:47equipment.
229. 8:49DeepSeek makes it perfectly clear that
230. 8:52this is a problem, that this philosophy
231. 8:54hasn't been solved yet, and what I've
232. 8:56done is perform something similar to the
233. 8:59DIS experiments documented in the
234. 9:01repository using a different model.
235. 9:04For this, I used Claude 3.1 Opus. I ran
236. 9:0860 messages per run and three runs per
237. 9:12design.
238. 9:13The same three rivals: the session with
239. 9:15the same neutral prompt, the session
240. 9:17that delegates first, and also using the
241. 9:19router with the workers.
242. 9:21What I achieved with DeepSeek, the
243. 9:23session that doesn't delegate was
244. 9:24actually worse.
245. 9:26Indeed, while working, messages
246. 9:28accumulate, and its slowest tasks took
247. 9:31about 60%longer.
248. 9:34What I didn't achieve with my experiment
249. 9:37is that, using the "delegate first"
250. 9:39expression, my single session delegated
251. 9:42in all three runs and tied with the
252. 9:45super-agent.
253. 9:46A phrase in the prompt and the result
254. 9:49was the same. Where did the super-agent
255. 9:52win? In context, it used almost three
256. 9:54times less than the "delegate first"
257. 9:56session.
258. 9:57This makes sense because that session
259. 9:59swallows the entire chat into a single
260. 10:01context that it doesn't let grow.
261. 10:05It also won in something that doesn't
262. 10:07show in the times, which is that in the
263. 10:08super-agent, each follow-up goes
264. 10:10directly to the worker for that task.
265. 10:13In the single session, it depends on the
266. 10:15main session remembering to pass it
267. 10:17along.
268. 10:18And my conclusion here is that the model
269. 10:20that obeys the phrase gives you almost
270. 10:22everything.
271. 10:23The architecture is useful when you
272. 10:25cannot trust it to obey. Or when the
273. 10:27chat grows so much that the context
274. 10:29becomes the problem. And these are just
275. 10:32three runs, so take it as a sign, not
276. 10:34really as proof. So, what would I be
277. 10:37adopting from this new DeepSeek
278. 10:39super-agent philosophy?
279. 10:42Well, before using or cloning this
280. 10:43repository and adopting it in production
281. 10:45, which is something I wouldn't do, I
282. 10:47would start with something much simpler
283. 10:48.
284. 10:50And that is, I would start by just
285. 10:52adding this message to one of my first
286. 10:55prompts when I begin a project or when I
287. 10:58am trying to achieve a clear objective.
288. 11:01And in my tests, for example, adding
289. 11:04something like every message comes to
290. 11:06you, so the work you do gets interrupted
291. 11:08.
292. 11:10From now on, prefer giving each job its
293. 11:13own sub-agent and you only do the small
294. 11:15stuff.
295. 11:17Copy the ideas, a session that
296. 11:19distributes a task by context and
297. 11:20filters before the model.
298. 11:23Its prototype is not yet a personal
299. 11:25machine, so I repeat, take it with great
300. 11:28caution.
301. 11:29In the genetic guild we build flows like
302. 11:32this all the time so your attention
303. 11:34yields much more than your agents. So
304. 11:36I'll leave the link for you to join
305. 11:38today in the description. And don't
306. 11:41forget: how many sessions do you have
307. 11:42open right now? I'll read you in the
308. 11:44comments if this video has been of any
309. 11:46use to you.
310. 11:47Please subscribe to this channel, turn
311. 11:49on the notification bell, leave a like,
312. 11:51share it so it reaches more people, and
313. 11:53I'll see you in the next one.
314. 11:55Yeah.