How One Fable Chat Manages 100+ Codex Agents — Transcript

Full transcript

1. 0:00Hey guys, I've
2. 0:02built out quite an interesting system
3. 0:04over the past few days that is just made
4. 0:07to be as simple as possible for having a
5. 0:11for having one orchestration agent
6. 0:14proactively push your tasks and goals
7. 0:18forward.
8. 0:19So this is based off of
9. 0:22the the agent that I use is like the
10. 0:23central orchestration agent. It is Fable
11. 0:27and I'm managing all of my tasks from
12. 0:30within notion but you could do it from
13. 0:33you know, I previously worked from
14. 0:35linear. I've worked from GitHub issues
15. 0:37which in fact was was sinking over to
16. 0:39linear originally when I I made this
17. 0:41system just
18. 0:43two days ago. It was with like markdown
19. 0:46files and it was kind of just like local
20. 0:48tickets, but I've decided to and I I am
21. 0:51actually finding that I do keep like
22. 0:52defaulting back to notion
23. 0:54and I think as like a kind of having a
24. 0:57central workspace for agents.
25. 1:00I do think it it it really works well
26. 1:02for that job. So I'll walk you through
27. 1:05like the kind of entire system just to
28. 1:07give you like a kind of high-level
29. 1:08overview and then I can give like a
30. 1:10little demonstration of it as well. I'll
31. 1:12try to keep this video shorter than than
32. 1:15all of the previous one. So
33. 1:17you have one main Claude chat and I'm
34. 1:21currently choosing to run on on Fable
35. 1:25and
36. 1:26the the main Claude chats it
37. 1:29chat it does not do any work itself.
38. 1:32It simply acts as the agent that you
39. 1:35speak to. It's like a chief of staff and
40. 1:37it
41. 1:39is just going to offload
42. 1:41all of that work to sub agents.
43. 1:45So you use the main Claude chat in this
44. 1:48case Fable for planning for talking kind
45. 1:51of brainstorming for discussing ideas
46. 1:54for clarifying the scope of work and
47. 1:56this is going to be the main agent and
48. 1:58it's I mean going in my opinion it's
49. 2:01it's the smartest agent
50. 2:03within the fleet. But again, you could
51. 2:05swap it out for any agent that you
52. 2:07prefer. It it it really doesn't matter.
53. 2:11And what it's going to do is it's going
54. 2:12to
55. 2:13offload all of the tasks, to delegate
56. 2:16all of the tasks to Codex,
57. 2:18to Codex app server. The reason I built
58. 2:20this system is I really hate speaking to
59. 2:225.6 soul. I in all honesty just find it
60. 2:26really difficult to work with. The
61. 2:28outcomes are good, but the
62. 2:31just the general kind of personality,
63. 2:34the
64. 2:35yeah, just the the the the general
65. 2:37experience of working with GPT 5.6 soul
66. 2:39I would say isn't a pleasant one, but it
67. 2:42kind of gets the job done. So, I'm using
68. 2:44Fable which is much easier to
69. 2:46communicate with. Like basically I was
70. 2:48just kind of finding myself getting a
71. 2:50little bit like annoyed and let's say
72. 2:52like more stressed than I would be when
73. 2:54dealing with GPT 5.6 soul. So, instead
74. 2:57I've put Fable there as like the middle
75. 2:59layer which also does all of the the
76. 3:02planning and architecture stuff as well
77. 3:04and GPT does as well. So, now you have
78. 3:06the benefit of two agents
79. 3:08that are able to interact with each
80. 3:09other which in my opinion it leads to
81. 3:12much higher token spend, but it also
82. 3:13leads to much better outcomes. Um so,
83. 3:16all of the actual work, the
84. 3:17implementation work is done by 5.6 soul.
85. 3:20Um again, it's using Open AI's official
86. 3:23Codex app server which just gives you a
87. 3:25adjacent API where you can kind of send
88. 3:28messages into a Codex chat. You can
89. 3:30archive chats, you can rename them, you
90. 3:32can
91. 3:33approve permissions, you can do anything
92. 3:35basically that's available within the
93. 3:37app um programmatically.
94. 3:40And then what I have is every 20
95. 3:43minutes, so again to go over the whole
96. 3:45architecture, every 20 minutes I have a
97. 3:47worker which is just like a cron job
98. 3:50that triggers a
99. 3:52Claude agent. It's not using Fable. I
100. 3:55think it was either Sonnet or maybe
101. 3:57Haiku, actually. And it lists our recent
102. 4:01chats in in Codex, and it just makes
103. 4:03sure that nothing's blocked, that
104. 4:05nothing needs our attention. So, it
105. 4:07makes sure that all work is still moving
106. 4:09forward, and then it will send that back
107. 4:11to our orchestration agent. So,
108. 4:14if there's something that needs our
109. 4:15attention, if there's something
110. 4:18Yeah, anything. The job of this
111. 4:20orchestration agent is to push work
112. 4:22going forward. And so, this worker just
113. 4:24lets the main agent here know, the chat
114. 4:26down here, that like something's
115. 4:28blocked. Maybe it needs a decision from
116. 4:30me. Maybe the orchestration agent can
117. 4:33solve it himself if we've previously
118. 4:34solved something similar and it's logged
119. 4:36in in decisions.
120. 4:38It it it kind Again, this is like how
121. 4:40I've built this system, and I'm kind of
122. 4:42comfortable with giving more autonomy to
123. 4:45Fable. But, you may not want to do that.
124. 4:47You may want to say like, you know, all
125. 4:49decisions have to have to have to be be
126. 4:51made by you. So, it's just a kind of
127. 4:53personal preference and kind of how much
128. 4:55you uh
129. 4:56you trust the agent.
130. 4:58Uh so, as I mentioned, like I'm working
131. 5:00from Notion tasks. I'll I'll show you
132. 5:02how that dashboard look looks in In
133. 5:04fact, I'll just show you right now just
134. 5:05to um
135. 5:08to give you an idea. So, what I'm
136. 5:09currently doing That should be priority
137. 5:11five. What I am currently doing then is
138. 5:16I'm just trying to see what ones I can
139. 5:18open. So, okay, I'm This is it's just a
140. 5:22simple like Notion Notion database.
141. 5:25Here's like a few like different tasks
142. 5:27I'm working on. I was just checking
143. 5:28there's nothing like confidential there.
144. 5:30So, this is the priority
145. 5:33tab. And so, in order just to not
146. 5:36overwhelm the Claude agent with like
147. 5:40tons of different things going on. I
148. 5:43mean, technically you could it can
149. 5:44probably handle it. It's more for my own
150. 5:46sanity, to be completely honest, but
151. 5:48I've limited it to only be working on
152. 5:50like five active tasks at one time. In
153. 5:53my backlog, I've got about 900 tasks,
154. 5:56but here I want a super clean view of
155. 5:59basically what is going on. Now, you can
156. 6:02also like click into it and each task
157. 6:04can have like its own
158. 6:07acceptance criteria and this like
159. 6:10progress circular icon thing like
160. 6:13automatically increases as different
161. 6:15tasks get
162. 6:17get ticked off. Agents then leave
163. 6:19comments. I'm not going to open these
164. 6:21because this is for it's hedge fund
165. 6:23stuff, so I just want to
166. 6:25to leave it off. In fact, this one's
167. 6:26fine. Like agents leave comments, so
168. 6:28like if you want to get like a quick
169. 6:30update on what's going on, they come in
170. 6:32every 30 minutes. It just gives a
171. 6:34summary, so you can kind of get an
172. 6:35overview of each task for what's going
173. 6:38on.
174. 6:38And on the far right-hand side, this is
175. 6:41the agent status.
176. 6:43So,
177. 6:44it it in progress means there is
178. 6:47actively activity happening that there
179. 6:49is an active chat working on this and
180. 6:52pushing it forward. If it's blocked, it
181. 6:54means it could need something from me.
182. 6:57You know, so needs plan, that will
183. 6:59automatically be kicked off. It will
184. 7:00start writing a plan and then it will
185. 7:02ask me for information. If it's done,
186. 7:04it's asking me if I want to approve the
187. 7:05spec. In progress, work is actively
188. 7:07moving forward. Blocked, something's
189. 7:09wrong. So, the Claude's chat will
190. 7:11automatically pick that up. I'll show
191. 7:12you that in a second. It's kind of a so
192. 7:15I don't think I really really need to
193. 7:16explain it. And then this on the last
194. 7:18side shows the last time that this chat
195. 7:22was successfully picked up. So, if it's
196. 7:25in progress here, it means that 19
197. 7:27minutes ago when the previous run went,
198. 7:30the worker that I mentioned that runs
199. 7:31every
200. 7:32every 20 minutes,
201. 7:34it said yeah, this is in progress and
202. 7:36it's moving forward. So, that should
203. 7:37update in in a minute as well. So, it
204. 7:39just lets let know
205. 7:41that this status kind of is up up to
206. 7:43date. And again, you could make that
207. 7:45work around more
208. 7:47more frequently or or so it down
209. 7:49depending on how many tokens you want to
210. 7:51spend.
211. 7:52In terms of like this system, I'm not
212. 7:55using Claude for any dev work at all.
213. 7:59And I've used up I in fact it has built
214. 8:02out some of this system, but it's not
215. 8:03really dev work. It's It's It's kind of
216. 8:05simple stuff. Mostly just like markdown
217. 8:07files and kind of some coordination
218. 8:10layers. And I've used up about 70% in
219. 8:14I think it was 3 days ago I started
220. 8:16building it when when whenever I
221. 8:17published that last video. Anyway, so
222. 8:19you're basically going to be paying
223. 8:20about $200 a month to Claude to just
224. 8:24have it as a communication layer,
225. 8:26which I'm personally fine with of course
226. 8:28for
227. 8:29I think a lot a lot of most people they
228. 8:32may not be happy with doing it, but for
229. 8:35me, you know, you It's like paying a
230. 8:37manager to kind of sit on top of the
231. 8:38team. And like why would you do that in
232. 8:40an organization? It's so that everything
233. 8:42doesn't come to you. So that there's a
234. 8:44bit of like a bridge between you. So
235. 8:46that you have the manager make sure that
236. 8:49work is happening. Like this is
237. 8:50literally like a micro manager. Every 20
238. 8:52minutes it's going and checking all of
239. 8:55the active work that I want to be
240. 8:57completed and making sure that it's it's
241. 8:59running.
242. 9:00So then it has hooks. So these are just
243. 9:03like kind of little prompts that fire on
244. 9:06different events. So turn end means when
245. 9:09Claude sends a response to me, it's the
246. 9:12end of his turn and it's going to launch
247. 9:15like two sub-agents essentially that
248. 9:18just go and say
249. 9:20you know, if if we message here saying
250. 9:22let's pause that task or let's start
251. 9:24that task or Fable tells me that task is
252. 9:27finished, it's going to read those
253. 9:29recent messages and it's going to go and
254. 9:31update the task in Notion as being
255. 9:33finished. Or if there's nothing to
256. 9:34update, it's not not going to do
257. 9:36anything. Yeah, it's also going to check
258. 9:38are there any blocked items and raise it
259. 9:41to the orchestrator. So, when we message
260. 9:43now, I showed you there was that blocked
261. 9:45task in in Notion, it should get picked
262. 9:47up automatically.
263. 9:48Let me also just say like this, this has
264. 9:51been built in 2 days. I'm getting value
265. 9:53out of it already, but it's like a work
266. 9:55in progress, but I think I will like
267. 9:57open source this um system just for for
268. 10:00you to go and like build on top of.
269. 10:02It may not even need to be open source,
270. 10:04but and you can just like
271. 10:06I don't know, copy this video like just
272. 10:07give a prompt to an agent to copy it and
273. 10:09it would be be very easy to do.
274. 10:12Um then also on session start, this
275. 10:15happens when you start a new Claude
276. 10:18session, but I was asking Claude some
277. 10:20other specifics and apparently it does
278. 10:22also happen if there's like a long
279. 10:24delay, it can also trigger a new session
280. 10:26start. I can't remember exactly what
281. 10:28what it said, but um
282. 10:30in that case, it's going to fetch all of
283. 10:33our our priorities
284. 10:35that are currently listed in Notion and
285. 10:38it's going to bring them locally and
286. 10:39then Fable knows that is the work that
287. 10:41it needs to push forward and it's also
288. 10:44going to check for any work that isn't
289. 10:45moving forward and it's going to
290. 10:48it's going to highlight that to me. So,
291. 10:50anything that is blocked that is moving
292. 10:52forward in the prompt, I've told it it
293. 10:54needs to fail like very loudly. It it's
294. 10:57not acceptable for work not to be moving
295. 11:00forward. If it's in that list of
296. 11:02priorities, it means and and the like
297. 11:05due date is today, it means this needs
298. 11:08to be pushed forward. This is a priority
299. 11:10for the organization, for the business,
300. 11:11this this need needs to get done. Then
301. 11:14as I've mentioned as well, there's the
302. 11:15watcher, sorry, there's actually two.
303. 11:17So, one of them is monitoring as I
304. 11:19mentioned already Codex app server.
305. 11:22This is checking for block chats for
306. 11:24maybe something times out, maybe
307. 11:26something goes wrong. I ran out of like
308. 11:28hard disk space earlier because I've got
309. 11:30so many work trees on on this machine
310. 11:34Um and so it like flagged that as an
311. 11:36issue. The other thing to say the thing
312. 11:40that I will say I do prefer about
313. 11:41Anthropic is their mobile app for
314. 11:46is their mobile app for Claude because
315. 11:48your
316. 11:50your Claude code chats are in sync with
317. 11:53with the mobile app and whereas that is
318. 11:55like an issue on
319. 11:57on GPT on the Codex app if you message
320. 12:00on the mobile app
321. 12:02it's not going to appear
322. 12:05in the same chat on the desktop. It just
323. 12:08exists but it just doesn't render so
324. 12:10it's like a front end thing. It's like a
325. 12:12minor thing but it is a little bit you
326. 12:15can kind of forget what where exactly
327. 12:16you left off. So it's just one thing
328. 12:18that I have noticed I do prefer and I
329. 12:20find like the thinking to be a bit
330. 12:22better and a bit kind of more stable on
331. 12:24Claude. So this entire system this
332. 12:26entire operation is made for kind of
333. 12:29being out and about and I know there's
334. 12:31loads of other people working on this
335. 12:32like remote control
336. 12:34problem as well. I was just playing with
337. 12:35another tool earlier today Orca for like
338. 12:38remote control orchestration. I've used
339. 12:40tmux for it as well. I do like their
340. 12:42solution.
341. 12:44Again kind of like what I said earlier
342. 12:47is I in the previous video is like I'm
343. 12:50just trying to keep things as complex as
344. 12:52necessary and as simple as possible and
345. 12:56it's like a battle that I'm kind of
346. 12:57fighting every single day. And with this
347. 13:00system
348. 13:01there is no external complexity. I'm
349. 13:04using Claude natively
350. 13:07and then I'm using the Claude app
351. 13:08natively. It can send me push
352. 13:10notifications on my phone. It does a
353. 13:12very good job of that. So if any work
354. 13:14gets blocked whilst this is operating on
355. 13:16on my local machine if anything goes
356. 13:19wrong here with any of these pull
357. 13:20requests I get a message in the official
358. 13:23in the official uh app and I can just
359. 13:26respond to it you know, if I'm out
360. 13:27walking, if I'm out for dinner, if I'm
361. 13:30you know, drinking coffee, like what
362. 13:31whatever I'm doing.
363. 13:32The work isn't blocked. This is just
364. 13:35acting as like a watcher, as a manager
365. 13:37of all of my Codex chats, of all of my
366. 13:39Codex sessions. Um
367. 13:42or again, whatever implementation agent
368. 13:43you want. It's monitoring all of your
369. 13:45agents 24/7 throughout the day. And you
370. 13:48have one of the smartest models do it
371. 13:50doing this for you. projects.md Sorry, I
372. 13:53will say as well, actually, I took the
373. 13:55original inspiration of this idea
374. 13:58from a guy that I saw he was an OpenAI
375. 14:02employee, in fact, and he had actually
376. 14:05open-sourced the system. It was a little
377. 14:07bit different to this, but it it was
378. 14:09like the general thesis with was the
379. 14:11same. And of course, he's working at
380. 14:12OpenAI, so he was pushing to to do this
381. 14:14with um using uh 5.6 Soul as Soul as
382. 14:19your
383. 14:20um chief of staff, whereas I'm instead
384. 14:24kind of saying I like 5.6 Soul for
385. 14:26implementation, but in terms of like the
386. 14:28chief of staff, I don't think it's very
387. 14:31good. Uh I I think you're much better
388. 14:33off using something like Fable, which is
389. 14:35just much nicer to communicate with. Uh
390. 14:37the other thing is
391. 14:40I think Claude is far, far better at
392. 14:44Let's say like sub-agent orchestration.
393. 14:47And so, by that, I mean natively within
394. 14:50Claude at launching loads of different I
395. 14:52mean, GPT has got better at it, but I do
396. 14:55prefer I I I just from what I've seen, I
397. 14:57do think that that Claude is actually
398. 14:59better at that. Uh so, Claude can kind
399. 15:01of spin up agents to kind of watch over
400. 15:03things, you know, if you've got really
401. 15:05high priority stuff, it can just put a
402. 15:06high core agent to go and watch um the
403. 15:09the implementation and make sure that
404. 15:11everything's kind of uh moving forward.
405. 15:13But anyway, yeah, projects is just my
406. 15:15active GitHub pro Sorry, this is the
407. 15:17folder structure as well, just so you
408. 15:18can kind of see what I'm in here. So,
409. 15:20the only thing that isn't represented
410. 15:22here is hooks because it's it's done
411. 15:23natively within
412. 15:25within Claude.
413. 15:27And sorry, just before I forget, I'll
414. 15:29also write out the
415. 15:32build skill skills directory. Okay. So,
416. 15:36yeah, so there's a projects.json,
417. 15:38which at the moment I'm working on about
418. 15:41five to eight things. I've got two
419. 15:43active projects.
420. 15:46And then I've got a cup just other side
421. 15:48projects that are just things that I
422. 15:49like to do for fun, things I find
423. 15:51interesting, you know, on a Saturday or
424. 15:53Sunday I'll often just like dedicate one
425. 15:55or two days a week to explore new ideas,
426. 15:57to try out new things. Some of them are
427. 15:59just like one one-time projects as well
428. 16:01that like just little tools that I kind
429. 16:02of build for myself. So, projects.json
430. 16:05just contains every project. It It tells
431. 16:09Claude if it's local only,
432. 16:11as in doesn't need to deploy on GitHub,
433. 16:13doesn't need to use work trees and all
434. 16:15of that, or is it just a local project.
435. 16:19And just kind of some things about like
436. 16:20my any kind of rules for working
437. 16:22specifically in that project that it
438. 16:24should be aware of. If the project is
439. 16:26like a big priority for me and and
440. 16:28that's kind of pretty much it. So, it's
441. 16:30just a very lightweight JSON. And then
442. 16:32forward slash projects
443. 16:34is a symlink, so it's just a it contains
444. 16:37all of the different projects that I'm
445. 16:39working on at the moment in subfolders.
446. 16:43And then decisions.md
447. 16:45I actually changed into a folder, so
448. 16:48it's decisions forward slash
449. 16:50decisions.md and then each project has
450. 16:55its own has its own decision log as
451. 16:58well. And the reason that I like to do
452. 17:01this is because if I have a preference
453. 17:03in doing something or I have
454. 17:05let's say I make an architecture
455. 17:07decision on like something like no,
456. 17:09don't don't build that with airtable
457. 17:12because we're in the process of
458. 17:13migrating to another database. If I give
459. 17:15a
460. 17:16a preferred approach of doing things
461. 17:18anything like that. The the rule is if
462. 17:21it's anything that the agent can't
463. 17:23understand from the code as in it's
464. 17:25being like verbal communication
465. 17:28that isn't in the code or in the doc
466. 17:30somewhere, it gets logged into or
467. 17:32anything that isn't obvious to the
468. 17:33agent.
469. 17:36It gets logged in a decisions.md folder.
470. 17:39I try I'm trying anyway to keep these
471. 17:41like as small and lean as possible just
472. 17:45to not blow the
473. 17:46the agent's context window every time he
474. 17:48opens this this file. It's it's the most
475. 17:51basic type of memory that that you can
476. 17:53have. I'll probably look at improving it
477. 17:55later on.
478. 17:56But yeah, it's just like the agents can
479. 17:58kind of look over it if they're not sure
480. 18:00what to do before coming to me with a
481. 18:01question. They can do a quick look over
482. 18:03the decisions and just see if we've
483. 18:05dealt with with something like this
484. 18:07before.
485. 18:09Then the next thing again, this is just
486. 18:11acting as like an orchestration layer.
487. 18:13It does no building itself. Everything
488. 18:15gets handed off to app server, but it
489. 18:17does do planning and kind of
490. 18:19clarification of all types of work. So
491. 18:22you can have kind of management skills.
492. 18:24So I've got I'm not even sure how many
493. 18:27skills I have, but one of the ones that
494. 18:28I made is the build skill
495. 18:31which I kind of mentioned in the last
496. 18:33video. This gets automatically invoked
497. 18:36whenever we're going to build something.
498. 18:38And it's basically a wrapper that sits
499. 18:41on top of the over superpowers skill
500. 18:45set. So it it runs brainstorming and
501. 18:47then it runs through implementation. The
502. 18:49only difference in this skill
503. 18:51again, it literally wraps and just says
504. 18:53call the brainstorming skill set. When
505. 18:56it gets to the
506. 18:59the specification part of the
507. 19:01superpowers workflow,
508. 19:03it goes through an adversarial audit
509. 19:07loop that I mentioned last time. So it's
510. 19:08going to spin up two sub agents. One of
511. 19:10them is going to be Opus 5, and one of
512. 19:13them is going to be GP T5.6 Soul.
513. 19:16With the Opus 5 agent, I then also get
514. 19:18Soul to actually order its work because
515. 19:21I
516. 19:22I do like Opus 5, but I just find that
517. 19:25if if you run its findings through GPT,
518. 19:29um GPT will kind of
519. 19:31find a load of things that Opus has a
520. 19:34tendency to over exaggerate. So, I use
521. 19:36Opus as a second opinion, but I I filter
522. 19:38it down anyway. So, anyway, one
523. 19:41GPT 5.6 Soul and then one
524. 19:44uh Opus 5 agent both audit the plan
525. 19:49and look for gaps. They look for
526. 19:50weaknesses. They look for things that
527. 19:52should be clarified, for things that
528. 19:53agents missed, for
529. 19:55um you know, dangerous behavior, for
530. 19:57code that could be like reused, or logic
531. 19:59that that
532. 20:01um that already exists that doesn't need
533. 20:03to be be rebuilt, all types of stuff. It
534. 20:05its job is to basically attack the plan,
535. 20:07and then it will come back and give the
536. 20:09orchestration agent here the um its
537. 20:12findings. The orchestration agent will
538. 20:14ask me, I'll clarify, and then it will
539. 20:16go back, and it will just loop over this
540. 20:18adversarial
541. 20:19review process. Now,
542. 20:21I mentioned this in the previous video,
543. 20:23but again, depending on what you're
544. 20:24building
545. 20:26is going to depend how many loops you
546. 20:28you you want to do with this audit
547. 20:30process because it can get expensive on
548. 20:32tokens, especially if you run 5.6 Soul
549. 20:35on on high reasoning. I've blown through
550. 20:38three uh pro accounts credit this week
551. 20:40because I've been building
552. 20:43changes into the actual trading system
553. 20:45that that I'm running, and of course,
554. 20:47it's if you're you're dealing with real
555. 20:49money, I much prefer to do like 10
556. 20:51adversarial iterations, but for
557. 20:53something small, you may only need to do
558. 20:56one, uh and for other stuff, you may do
559. 20:58three. So, again, it's going to be like
560. 20:59personal kind of preference.
561. 21:01And so, yeah, I that is like the build
562. 21:03skill. The other thing as well is um
563. 21:07what I was finding is that often the
564. 21:09agents could default into using
565. 21:12superpowers.
566. 21:13And on some of the work that I do and so
567. 21:16it would try to like build everything
568. 21:17and superpowers is in itself a bit of a
569. 21:21like I love it. It's my favorite set of
570. 21:23skills but it can be a little bit of
571. 21:24like overkill or too much of an
572. 21:26elaborate workflow for like really small
573. 21:28changes. Like I was asked to put in like
574. 21:30a button and that makes one change in
575. 21:32the database and it went through this
576. 21:34whole superpowers workflow and it took
577. 21:35about
578. 21:36I don't know. It was about 3 hours for
579. 21:38something that that
580. 21:40really was like it just wasn't
581. 21:41necessary. It was complete overkill.
582. 21:44So that's one thing. So the build skill
583. 21:45as well also audits the task and tries
584. 21:49to understand is this something worth
585. 21:51using superpowers on or is this just
586. 21:53like a really quick fix that can just be
587. 21:55validated with like a test or two.
588. 21:58And then also on some of the work that
589. 22:00I'm doing as well, it can be
590. 22:02you know, about auditing stuff and I had
591. 22:06cases where it was trying to build a
592. 22:07solution before we had even actually
593. 22:09audited what what what the actual
594. 22:11problem was like investigated what what
595. 22:13the problem was. Like it would be more
596. 22:14like you know, things going wrong with
597. 22:17within the the fund that's in like
598. 22:19something isn't operating properly. Go
599. 22:21and investigate what's going wrong and
600. 22:22then let's actually harden that that
601. 22:24system. So this kind of would choose
602. 22:28what type of agent does it route to?
603. 22:31Does it launch an investigation first?
604. 22:33Does it go straight into like a
605. 22:34superpowers workflow? Does it just do
606. 22:36like a
607. 22:37small fix? And so that was another thing
608. 22:40there. I've got one or two other skills
609. 22:42but honestly off the top of my head I
610. 22:44completely forgotten what what they are.
611. 22:46Anyway, I'll give you an example of like
612. 22:47the system in action. So if we come back
613. 22:50into
614. 22:51if we come back into the tasks here
615. 22:54I don't know why my shortcut's not not
616. 22:57working there.
617. 22:59Anyway,
618. 23:01right. So we've got a few things that
619. 23:02are blocked.
620. 23:04So, what we should be able to do here is
621. 23:05if we come we've got fable on I can
622. 23:07literally just say hi. So, I shouldn't
623. 23:09need to prompt. Yeah, okay, right. So,
624. 23:12without even saying this is what I mean
625. 23:14about I just wanted this system to
626. 23:15basically be
627. 23:17as proactive as possible. So, it's gone
628. 23:21and seen
629. 23:24that uh
630. 23:251554
631. 23:27on notion task 1554 which is this needs
632. 23:31a user decision. So, it works in
633. 23:32priority order by the way.
634. 23:35Now,
635. 23:36what I I I I didn't like it before kind
636. 23:39of overwhelmingly with like a ton of
637. 23:41information. So, look now it's looking
638. 23:43through like all of the local files.
639. 23:46It's going through the projects and the
640. 23:48ledger of work which is the local
641. 23:50history of things that we've built
642. 23:52just to kind of check like where we got
643. 23:54on to to to kind of check its notes. Um
644. 23:57and
645. 23:58again, just
646. 24:00the whole thing of basically telling
647. 24:03this agent like
648. 24:05you do not build anything. You apply all
649. 24:09of your intelligence simply into pushing
650. 24:11the project forward. Also, sorry. It's
651. 24:14kind of very similar to the concept that
652. 24:16open claw and Hermes or MS
653. 24:20agent were kind of built on top of of
654. 24:23being like a proactive personal
655. 24:25assistant. So, it's kind of built off of
656. 24:26that, but this is more for like a chief
657. 24:28of staff.
658. 24:30I I've I did previously try a system
659. 24:32like this with Hermes agent and just
660. 24:34wasn't able to um
661. 24:36to to get it to work at this same level.
662. 24:38So, again, yeah, it's just going to push
663. 24:39the work forward. It's just saying,
664. 24:41"Look, this is blocked clean mergeable."
665. 24:43The one thing that I really hate is when
666. 24:45it sends me these like super long
667. 24:47messages. So,
668. 24:48uh
669. 24:49I'm still trying to like perfect the
670. 24:52communication again because
671. 24:54I don't like how long is it going to
672. 24:55take me to read this many I I don't
673. 24:57know. Maybe like 30 seconds or 40
674. 24:59seconds, but if it's just more concise,
675. 25:02if it's more clear and concise, again,
676. 25:04you if if you're dealing with hundreds
677. 25:06of these messages per day, if you can
678. 25:08shave off like 30 seconds off of
679. 25:11hundreds of messages per day or even
680. 25:13more than that, again, just on the
681. 25:15amount of emotional bandwidth that you
682. 25:17need to apply to each message, if it can
683. 25:19just Yeah, exactly. Merge PR, check
684. 25:21screen.
685. 25:22Um
686. 25:24Okay, so now he's going to go and tell
687. 25:26that agent, again, this will be a Codex
688. 25:28agent that's done the work. It's going
689. 25:29to send that API request over to back
690. 25:32back to the Codex, sorry. So, this one,
691. 25:34it's needs user decision, so that should
692. 25:36now become unblocked.
693. 25:38It it needs to send the message and then
694. 25:40the hook runs after he's sent his final
695. 25:42message, which needs another agent to
696. 25:43run. So, there's going to be like a
697. 25:46maybe like a 30-second, maybe like a
698. 25:481-minute delay after it sends its final
699. 25:51turn, because that's when the hook runs.
700. 25:53So, it's not like real time. You could
701. 25:56like
702. 25:57the reason that I'm also using hooks is
703. 25:59because again it uses Haiku. So, number
704. 26:01one, you save on context, but
705. 26:03that is all all of the kind of um
706. 26:07system side of things is happening on
707. 26:09the back end without this Fable agent
708. 26:11needing to even think about it, because
709. 26:13otherwise he can forget about stuff,
710. 26:15part of his kind of context, which I I'm
711. 26:18like obsessed with preserving these
712. 26:21agents' context as as much as possible.
713. 26:25is going to be used up on kind of
714. 26:27maintaining the system, which is exactly
715. 26:28what we don't want. This Fable agent's
716. 26:30only job is to work with me on pushing
717. 26:34all all all all all all of the the work
718. 26:36for And previously, when I was doing
719. 26:38this, I would either have
720. 26:41about 30
721. 26:43terminals open running individual
722. 26:46sessions
723. 26:48or
724. 26:50running natively within the Codex local
725. 26:53app. Uh because there have been like a
726. 26:56lot of improvements in there recently
727. 26:58that I do like and I have have been
728. 27:00using, but this is just such is just a
729. 27:03much more peaceful way of working, I
730. 27:05think is the easiest way of describing
731. 27:07it where again it's costing like $200 a
732. 27:10month to um
733. 27:12just to have this the this level of
734. 27:14intelligence that sits in between you
735. 27:18as as the user. So, it just reduces the
736. 27:22amount of like nonsense that that you
737. 27:24have to deal with. It refines down all
738. 27:26of the chats and I think that it's a I'm
739. 27:29biased because I built it, but I I think
740. 27:31that this is like the cleanest um
741. 27:35the the cleanest way of working that
742. 27:37I've come across for manage for for
743. 27:39building a ton of stuff. So, like
744. 27:41there's a lot of people working on
745. 27:42building out these like software
746. 27:44factories
747. 27:46at at the moment and I think that this
748. 27:48is like a really interesting way of
749. 27:50doing it. Again, I'm not the the person
750. 27:52that has kind of come up with this idea.
751. 27:54I've just taken a a few different things
752. 27:56that I I saw this open AI guy the other
753. 27:58day who had like a kind of sim- similar
754. 28:01thing of having a chief of staff and
755. 28:04having one subfolder with all of your
756. 28:06projects. And and I thought it was so
757. 28:08interesting to come from a guy who works
758. 28:10at open AI to say, "Don't initiate Codex
759. 28:13within each project like you normally
760. 28:15would. Just put him one layer above and
761. 28:17he's going to know what what what what
762. 28:19to go into." Which I was quite surprised
763. 28:21to hear hear them say actually. So, I
764. 28:24thought like why not? That does make
765. 28:25sense especially for for myself working
766. 28:28on multiple projects. Let's
767. 28:30Let's give it a try. So, anyway,
768. 28:31basically now I just sit here all day
769. 28:33just working with like one
770. 28:37one chat here. You you could have like
771. 28:39multiple going on and like as always the
772. 28:43the problem that you you still have is
773. 28:45like you know, okay, the problem that
774. 28:47you have is like your
775. 28:49how much you're willing to pay on AI
776. 28:51subscriptions.
777. 28:53I personally get tremendous value out of
778. 28:56the subscription, so I'm happy to take
779. 28:58another subscription. So, my limiting
780. 29:00factor in terms of scale
781. 29:03it is down to
782. 29:04how
783. 29:06how many decisions I can make per day.
784. 29:09Like the one after that, I guess would
785. 29:11be like how much time I have per day,
786. 29:13but really it comes down to like how how
787. 29:15many
788. 29:16good decisions I can make per day before
789. 29:19I I get very fatigued, and I'm I'm
790. 29:21really quite obsessed with this at the
791. 29:25moment. But again, with this system, I
792. 29:26can go out on on my phone, and it it's
793. 29:29just going to ping me whenever it gets
794. 29:30stuck. So, it knows what the priorities
795. 29:32are. It knows
796. 29:34what to push forward. I've already done
797. 29:35the planning work. If there's no plan,
798. 29:37it's going to proactively ask me and
799. 29:40start making it, and then it's going to
800. 29:41commit the plan. And once it's got the
801. 29:43plan, it can kind of go go ahead with
802. 29:45the with with the jobs. The other reason
803. 29:48I built this as well, I don't know if
804. 29:50anyone else has experienced this, but
805. 29:51when working with Superpowers
806. 29:53brainstorming, there was like a really
807. 29:55big architecture refactor that I'm still
808. 29:58doing that I mentioned already on the
809. 30:01on the trading system to move it into
810. 30:03being a queue-based system in instead of
811. 30:05to of of sending orders at real time. It
812. 30:08just kind of makes more sense. Anyway,
813. 30:10it's a huge refactor, and it it involved
814. 30:13it's about eight different pull
815. 30:14requests, and it was it's basically just
816. 30:17too much work for Superpowers to handle
817. 30:21because it it's it's like a huge
818. 30:23project, and so it kept stalling. And
819. 30:25so, I wrote that plan
820. 30:28about a week ago, and it's still not
821. 30:30implemented with
822. 30:31these agents. So, that's why originally
823. 30:34I because they kept stalling because
824. 30:36like they would have to compact their
825. 30:37context so much. So, that's why I got
826. 30:40this watcher agent in place in the first
827. 30:42place, and then I thought, "Okay, this
828. 30:44makes complete sense. Let me actually
829. 30:45just build a um
830. 30:47a system around this." So, yeah, I will
831. 30:49upload the I'll put this on GitHub. If
832. 30:52it's interesting for you or just to play
833. 30:54around with it, by all means, you you
834. 30:56can use it for free, or you can even
835. 30:57just take a screenshot of this and send
836. 30:59it to your agent and kind of customize
837. 31:01it.
838. 31:02I think it's very good. It's very, very
839. 31:04flexible. It's very simple. So, there's
840. 31:06kind of really a limit. Like, literally,
841. 31:08all we're talking about is send API
842. 31:10requests to
843. 31:12um to OpenAI agent, have a tasks
844. 31:16database. So, again, that could be
845. 31:17GitHub, it can be Notion, it can be
846. 31:19Linear, what whatever you prefer to use.
847. 31:21Move all of your projects in into one
848. 31:23subfolder, which is where this Chief of
849. 31:25Staff agent works from, and then we're
850. 31:28literally just talking about a couple of
851. 31:29hooks
852. 31:31that basically monitor the work that's
853. 31:32going on and make sure that it keeps
854. 31:34getting pushed
855. 31:36pushed forward. So, um yeah, anyway, I
856. 31:38hope this was helpful. Thank you so much
857. 31:40for watching. Any questions, let me know
858. 31:42in the comments.