Tmux + Fable = Cut 35% less token — Transcript

Full transcript

1. 0:00You probably experienced your cloud
2. 0:01quota burning out Fable quota so fast
3. 0:03that you hit a limit very quickly. And
4. 0:05that's probably because you are not
5. 0:07using the model right. So, as we get
6. 0:09stronger and stronger frontier model
7. 0:10like Fable 5 or GPT-5.6 so max, even
8. 0:14though those model are extremely
9. 0:15powerful, but also they are expensive
10. 0:18and slow to run. That's why more and
11. 0:19more people, even Cloud Code official,
12. 0:22are recommending this practice that many
13. 0:24use Fable 5 as the advisor or planner,
14. 0:27but use smaller model like Sonnet 5 to
15. 0:29be the executor. Cuz even though Sonnet
16. 0:315 price is only 20% of Fable 5, but its
17. 0:34performance is almost equivalent to Opus
18. 0:374.8, which is strongest model a few
19. 0:40months ago. And there are two ways Cloud
20. 0:41Code official has shown. One is that
21. 0:43using Sonnet 5 as the main executor
22. 0:46agent, but only do it to call to get
23. 0:48Fable 5 as the advisor to almost review
24. 0:50its plan and send advice. Or second
25. 0:52method is that you use Fable 5 as the
26. 0:55main agent where it will make a plan and
27. 0:57spin up worker agent using smaller model
28. 1:00for execution. And even Devin is
29. 1:02introducing this new harness type called
30. 1:04Devin Fusion that claim to achieve
31. 1:06similar or even better performance than
32. 1:08Fable 5, but with 35% lower the cost.
33. 1:12And fundamentally, their Fusion harness
34. 1:15is just Cloud Code's Fable 5
35. 1:16orchestrator plus Sonnet 5 worker model.
36. 1:19And Devin actually compare the two
37. 1:21methods Cloud Code mentioned. One is the
38. 1:24Sonnet 5 as the main executor, but use
39. 1:26Fable 5 as advisor versus using Fable 5
40. 1:29as main orchestrator, but spin up a
41. 1:31worker agent using smaller model. And
42. 1:34the recommendation from them is actually
43. 1:35very clear that the Fable 5 orchestrator
44. 1:38model is working a lot better because
45. 1:40with advisor model, you actually need to
46. 1:42pay a very expensive price because
47. 1:45advisor model will need to read the full
48. 1:47conversation history from the main
49. 1:49agent, rather than just trigger a cached
50. 1:51context. And cached token is about only
51. 1:5410% cost of new input token. So, turn
52. 1:57out using this model is actually going
53. 1:59to work a lot better and cost-efficient.
54. 2:01And to make it work well, the main ideas
55. 2:04here is that each agent here should not
56. 2:06be just a basic sub agent, but instead
57. 2:09Devin is calling a sidekick. And what
58. 2:10this means is traditionally, when you
59. 2:12use Cloud Code or Code X and start a sub
60. 2:15agent, it generally just spin up a new
61. 2:17agent session on the fly. And once agent
62. 2:20finish, the last message will be sent
63. 2:22back to the main agent. But, let's say
64. 2:24main agent realize there's something
65. 2:26wrong with it. And when you want to make
66. 2:28edits, the new task tool will basically
67. 2:30spin up a new sub agent session that has
68. 2:32no context of what has been discussed
69. 2:34before, which means a lot of wasted
70. 2:36token on rewriting the same thing. But,
71. 2:38with sidekick model that Devin
72. 2:40mentioned, the key difference here is a
73. 2:42sub agent actually is a persistent
74. 2:44session, which means if the main agent
75. 2:46has some feedback, it will just send
76. 2:48back a follow-up message to the same sub
77. 2:51agent, which will inherit all the past
78. 2:53context. And those past context, since
79. 2:55we're being the cached token, it's going
80. 2:57to be extremely cheap, yet very
81. 2:59powerful, because it has a full context.
82. 3:01And if you're using Cloud Code, you're
83. 3:02probably already pretty familiar with
84. 3:04this model, because that's literally
85. 3:06what agent team means.
86. 3:12You can literally just prompt Cloud Code
87. 3:14to use agent team to do certain task.
88. 3:17Instead of just starting a one-time sub
89. 3:19agent, you can run this persistent
90. 3:21session. Once it's finished, the main
91. 3:23agent will be notified, and it will be
92. 3:25continuously sent a follow-up message to
93. 3:27this persistent agent session, instead
94. 3:30of starting a new session. Cuz the new
95. 3:32Cloud Code version already have the send
96. 3:33message tool that allow you to send a
97. 3:36new message to an existing agent
98. 3:37session. And once those agent finish, it
99. 3:39will stay resumable to receive those
100. 3:42message until it calls a shutdown
101. 3:44request. And when Cloud Code starting a
102. 3:46agent team session, it can even define
103. 3:48the model as well. So, So, you're using
104. 3:50Cloud Code, it's actually very easy for
105. 3:52you to achieve this type of Fable 5S
106. 3:54main coordinator and planner and Sonnet
107. 3:575 or even smaller model as executor. And
108. 4:00you can achieve this workflow by simply
109. 4:02adding this delegation rules into your
110. 4:04cloud.md file. It's basically saying
111. 4:06that you are the coordinator, that will
112. 4:08be design, design, and review. And we
113. 4:09delegate hands-on execution to the
114. 4:11executor sub-agents using the Sonnet
115. 4:13model. It lists out what are things that
116. 4:15should be delegated and what are things
117. 4:17that should be kept in the coordinator
118. 4:19main agent. Which are specific things
119. 4:21like design, planning, architecture, or
120. 4:24just tiny edits. And there's one section
121. 4:26that's pretty important about road
122. 4:27check. So, the sub-agent would know it a
123. 4:30sub-agent, so it don't keep spinning up
124. 4:32more nested agents. So, with this one, I
125. 4:34can simply start a session asking to
126. 4:37help me design and build a to-do app. It
127. 4:39will ask me question, do some planning,
128. 4:41and typically for complex task, it will
129. 4:44create a frozen spec in the task folder,
130. 4:46then make the plan very clear, so
131. 4:48there's no ambiguity for executor. Then
132. 4:51spin up a sub-agent with Sonnet 5 to do
133. 4:54the actual execution. And the sub-agent
134. 4:56will receive this task saying the role
135. 4:58is executor, do the work themselves,
136. 5:00don't spawn sub-agents, and then read
137. 5:02the spec first, and do the
138. 5:03implementation. And because it is Sonnet
139. 5:055, it's actually going to be much faster
140. 5:08than you implementing with the Fable 5.
141. 5:10And I have included my delegation rules
142. 5:12in the AI Builder Club GitHub repo. So,
143. 5:14you can copy and insert into your own
144. 5:16cloud.md. If you're interested in
145. 5:18learning more, there are more
146. 5:19step-by-step tutorial on this coding
147. 5:21agent workflow, and we have weekly live
148. 5:23workshop to cover those. I have included
149. 5:25link in the description below, so you
150. 5:26can come and join if you're interested.
151. 5:28However, this only works if you're using
152. 5:30Cloud Code. But in reality, probably you
153. 5:32have a few different AI coding agent
154. 5:34that you're using. Some of you probably
155. 5:36using Gemini CRI because the Gemini
156. 5:38model is still one of the best and
157. 5:40cost-effective model for front-end
158. 5:42design, or the recent Kimi case 3 model.
159. 5:44So, in an ideal world, you still want to
160. 5:46use the best possible model and harness
161. 5:49as a coordinator, which can be the Cloud
162. 5:51Code or Codex using the GPT- 5.6 So Max.
163. 5:55But the actual worker you should be able
164. 5:57to free to use any of those based on
165. 5:59your preference. And Codex team has been
166. 6:01pretty sneaky. They introduced this
167. 6:03Codex plugin CC tool, which is a Cloud
168. 6:06Code plugin that allow you to delegate
169. 6:08task to a Codex agent. So you just
170. 6:11install this plugin and add the OpenAI
171. 6:13Codex, then reload plugins. And now I
172. 6:16can prompt it to delegate to Codex agent
173. 6:19to help improve our UI of to-do app on
174. 6:22this URL, which is first version that
175. 6:25our Cloud Code agent did. And then it
176. 6:27will start this Codex session. And here
177. 6:29is a bit of weird cuz it basically start
178. 6:31a sub-agent to call this Codex agent.
179. 6:33It's probably because our
180. 6:35coordinator/executor
181. 6:37setup in cloud.md. But while it is
182. 6:40running, you can see this Codex plugin
183. 6:42giving us a few command. One is Codex
184. 6:45rescue. This basically treat Codex as a
185. 6:48sub-agent. Or a Codex review, which
186. 6:50allow you to get a Codex to review the
187. 6:52local git. And Codex result allow Cloud
188. 6:55Code to fetch the output from a specific
189. 6:57Codex session. Same as Codex cancel. And
190. 7:00it can even do Codex transfer, which
191. 7:02should basically take the conversation
192. 7:04history to this point but fork over on
193. 7:07Codex side. Now if I check my to-do app
194. 7:09again, you can see that it update the
195. 7:11UI, which looks probably more OpenAI-ish
196. 7:15while the Entropic feeling. The one
197. 7:17limitation with this Codex plugin is
198. 7:19that I don't think there's a way for you
199. 7:21to keep sending message to the same
200. 7:23Codex session. If here I ask it to, "Can
201. 7:25you send a follow-up message to this
202. 7:27Codex session to say good job?" Oh,
203. 7:30actually I am wrong. It actually can
204. 7:32send back message cuz there's a resume
205. 7:35command, which will allow it to send
206. 7:37message to the existing session. So it
207. 7:40can actually achieve that by doing this
208. 7:41hack. So, the bridge between CodeX and
209. 7:44Claude Code is pretty much there. But,
210. 7:46what if you wanted to actually talk to
211. 7:48any other coding agent that you are
212. 7:50using like Gemini CLI, Guark Build, Pi
213. 7:53Agent, Open Claw, any others. This is
214. 7:56actually a way by using Tmux. So, if you
215. 7:58don't know, Tmux is basically a terminal
216. 8:00multiplexer. It allows you to start
217. 8:02multiple different persist terminal
218. 8:04sessions and continuously read and
219. 8:06manipulate any other terminals. For
220. 8:08example, I can run this Tmux split
221. 8:10window -h, which means it will start a
222. 8:12new terminal session on the right side
223. 8:14of the screen. And inside that terminal
224. 8:16session, I want to run Claude Arsay and
225. 8:18with this message. It will spin up a new
226. 8:20session on the right and send out this
227. 8:22message. And what's cool is that on one
228. 8:24hand, this session is persist. You can
229. 8:26continuously sending more message to it.
230. 8:29For example, I can do Tmux send keys t,
231. 8:32which is the target, and dollar one
232. 8:34means the second terminal sessions on
233. 8:37the screen, "Tell me a joke." And then,
234. 8:39I'll do Tmux send keys enter. This one
235. 8:43will basically control the terminal on
236. 8:45the right side to typing this message
237. 8:47and click enter to send out message. And
238. 8:50once I run that, you can see on the
239. 8:52right side, a new message has been sent.
240. 8:54And I can also run this Tmux capture
241. 8:57pane -p -t .1. So, this means it will
242. 9:02actually capture what's showing on that
243. 9:04terminal and print out the result. So,
244. 9:06if I click enter that, then it will read
245. 9:09the result from the screen on the right
246. 9:12side and return that result in the
247. 9:14terminal. So, with this one, you
248. 9:16probably start getting the idea that you
249. 9:18can just tell your agent to use Tmux to
250. 9:20start new agent session just like you as
251. 9:23human. And through that way, it can
252. 9:25technically control any coding agent and
253. 9:27use that as a real agent teams. The only
254. 9:30caveat here is that it's easy for you to
255. 9:32start new session in this way. But, how
256. 9:34can the the agent actually notify back
257. 9:36the main agent that it has finished task
258. 9:39or it needs some help. So, there are
259. 9:41some hacks you can do. So, in tmux,
260. 9:43there's a way for each terminal to send
261. 9:46signals like
262. 9:47by doing this wait for -s with a
263. 9:50message. So, what do you technically can
264. 9:52do is start a tmux session and instruct
265. 9:56agent that once it's finished send this
266. 9:58special message. And then run this tmux
267. 10:01wait for this special message. You can
268. 10:03even just pipe and printing out the
269. 10:05section after it finish. So, if I do
270. 10:07that, it will restart the new agent
271. 10:09session and once it finish, it will run
272. 10:11this tmux wait for down hello world page
273. 10:14command so that this command will
274. 10:16finish. And you can imagine you can
275. 10:18basically instruct your agent to follow
276. 10:19this pattern, run background command so
277. 10:22that once it finish, it will be
278. 10:23notified. And I actually created a skill
279. 10:25called open agent teams. It's basically
280. 10:28encapsulating this tmux hack alongside
281. 10:30reference and adapter for each harness
282. 10:33like cloud code, code ask, grok pi, open
283. 10:35code and maybe others you can add to
284. 10:37achieve those kind of agent team
285. 10:39experience for any agent. And it
286. 10:41actually kind of works in any terminal
287. 10:43you're using. Like I can just open a
288. 10:45normal terminal and start grok and just
289. 10:47give a command. Ask code ask to make a
290. 10:49joke about engineer who write code and
291. 10:51after finish delegate to pi agent to
292. 10:54review the joke and also cloud code high
293. 10:57cool model to check the grammar in
294. 10:59parallel. And after all finish, you do
295. 11:01the final review and tell me. Right, so
296. 11:03you can do things like this. And with
297. 11:05this skill, I actually encapsulated the
298. 11:07short script that is encapsulating all
299. 11:09those tmux kind of practice but exposing
300. 11:12like start a new agent session, send
301. 11:15message to a session, wait for the
302. 11:17result, see the result, things like
303. 11:19that. So, you can see if firstly run
304. 11:22this command, start a code ask session
305. 11:24and this is kind of ID with a prompt,
306. 11:26then it run this peak command to quickly
307. 11:29check if this session actually running
308. 11:31well Cuz sometimes those coding agent
309. 11:33session can have those kind of
310. 11:35permission confirmation screen. And in
311. 11:38this case, it is working well, then it
312. 11:40will run this command to wait for the
313. 11:42session to finish with a max timeout.
314. 11:44And the session is running in the
315. 11:45background. And once finished, it will
316. 11:47do this result check. And here you can
317. 11:50see that the Codex actually wrote a
318. 11:52summary, but the summary didn't include
319. 11:54joke itself. So, then it can send a new
320. 11:57message to a specific session to ask to
321. 12:00print out the result. Then it got a
322. 12:02joke, stopped the Codex session, start a
323. 12:04pie review session for this specific
324. 12:07joke. And meanwhile, also start a haiku
325. 12:09grammar session with a special prompt.
326. 12:12And you can see there's some issues that
327. 12:13is blocked by the classifier. So,
328. 12:15there's probably some improvements that
329. 12:17we can do for the script as well, but it
330. 12:19is able to just self-healing and wait
331. 12:21for the result from both this pie review
332. 12:23and haiku grammar check. And also see
333. 12:26that the haiku grammar session is in
334. 12:28this kind of interactive mode. Then it
335. 12:30was able to just do the enter to control
336. 12:33this terminal session. And in the end,
337. 12:35return back the final version to me. And
338. 12:38I have included this open agent team
339. 12:40skill inside the AI Build Club repo
340. 12:42here. So, you can just copy and use. It
341. 12:46has this skill file talking about how to
342. 12:48use this script and also a reference
343. 12:50file for the cloud.md or agent.md file
344. 12:53that including this delegation rule
345. 12:55section. So, you can just copy paste
346. 12:57over. But meanwhile, if you really want
347. 12:59kind of premium experience, there are
348. 13:01also two like Herd and Orca that has
349. 13:04those orchestration CLI feature built in
350. 13:07and pair with a kind of UI layer. So,
351. 13:09it's pretty easy for you to see the
352. 13:10result. And I test almost all those
353. 13:13options. Personally, I found Orca is
354. 13:15providing the best experience. So, Orca
355. 13:17has this orchestration skills, which is
356. 13:20a CLI plus Q come out of box. And you
357. 13:23can basically install it. And once you
358. 13:24install that, it is basically achieving
359. 13:26the similar open agent delegations in
360. 13:29there I have to here. But because it's
361. 13:31very tied to the interface they have
362. 13:33here, the experience became very
363. 13:35intuitive. I'll give you one example.
364. 13:37Let's say I just start a kind of new
365. 13:40report, and in this report we don't have
366. 13:42any uh skills. It just have that
367. 13:44orchestration skill come out of box. I
368. 13:46can basically tell the same thing, like
369. 13:48help me ask Codex to make a joke about
370. 13:50any new whoop code. After finish
371. 13:52delegate to Pi, except I will also add
372. 13:55the skill to just make it more explicit.
373. 13:57Sometimes they will just do it by itself
374. 13:59as well. But just for demo purpose, if I
375. 14:02enter this, it will use the skill which
376. 14:04will invoke Orcas kind of still I that
377. 14:07they build in. And then first they start
378. 14:09a Codex session. But the really cool
379. 14:11thing here is that you can see this new
380. 14:14session just pop up on the right side.
381. 14:16So, it's actually very easy for me to
382. 14:18see all the sessions end-to-end, and I
383. 14:20can just put the sub-agent session on
384. 14:23the same screen here. Uh and on the left
385. 14:25side, I just have this kind of hierarchy
386. 14:28view to show me what are the
387. 14:30sub-sessions under the session. The same
388. 14:32thing, it also start this Pi agent for
389. 14:34the review, as well as this Haiku agent
390. 14:37for the grammar check. And again,
391. 14:39everything is just very well tracked in
392. 14:42their interface. I can just go into any
393. 14:44sub-agent to continue the conversation
394. 14:46there, and uh once I finish it will like
395. 14:49automatically communicate back to the
396. 14:51main agent. And I believe they will even
397. 14:53have some kind of special data
398. 14:55visualization to help you see how the
399. 14:57tasks orchestrated across different
400. 14:59agents. So, if you want a more kind of
401. 15:01packaged experience, I highly recommend
402. 15:03you check out Orcas well. They do have
403. 15:06so many good stuff built in, like the
404. 15:08Orca mobile experience, the Kanban view
405. 15:10out of box, the token usage tracking,
406. 15:13all the good stuff. And it's totally
407. 15:14open source and free, and they're
408. 15:16growing like really, really fast. For
409. 15:18the past few weeks, it has been my daily
410. 15:20driver for communicating with agents for
411. 15:22the past few weeks. And we actually did
412. 15:24a 1-hour deep dive workshop on some of
413. 15:26advanced AI coding workflow with Alka
414. 15:28founder. We actually showcase
415. 15:29step-by-step of how they were managing
416. 15:31to ship hundreds of commits to the Alka
417. 15:34product itself. In Octrade advanced
418. 15:36agent workflow across Grok, Codex, Cloud
419. 15:39Code, Open Code. You can also join AI
420. 15:40Build Club to see the in-depth workshop.
421. 15:42So, I hope you enjoyed this video. As
422. 15:44mentioned, feel free to get the skills
423. 15:46that I put in for either using this open
424. 15:48agent team skill as well as this
425. 15:50Contagion Rules section for your
426. 15:53cloud.md or agent.md file to achieve
427. 15:55those kind of Octrader worker setup. So,
428. 15:57that's pretty much for it today. Thank
429. 15:59you and I see you next time.