AI Software Factory z Grok Bot i Claude — Transcript

Full transcript

1. 0:00Imagine your agents organizing their own
2. 0:02work, testing it, and letting you know
3. 0:04when everything is ready. And all you do
4. 0:06is record a voice memo on your phone
5. 0:08telling them what to work on next.
6. 0:10Sounds cool, right? That is exactly what
7. 0:12my AI workflow has looked like for the
8. 0:14past few weeks. The whole thing is based
9. 0:16on Grokbot, which I use to manage all of
10. 0:18my projects. Today, I will show you how
11. 0:21I build my own private software factory
12. 0:23by combining several tools and
13. 0:25approaches in a simple way.
14. 0:28We will go through each of them so you
15. 0:29can get the best out of it and tailor it
16. 0:31to your own needs.
17. 0:33Let’s go. Alright, but do we even need
18. 0:35this?
19. 0:37Isn't Cloud Code or code on our own
20. 0:39computer enough?
21. 0:41I believe the era of agents living on
22. 0:43our laptops is slowly coming to an end,
23. 0:45and soon we will migrate most of our
24. 0:47work to the cloud.
25. 0:49Locally, we are limited by our hardware
26. 0:51power and the fact that it must stay
27. 0:53turned on all the time.
28. 0:55Plus, there’s the inability to work on
29. 0:58multiple devices and security concerns
30. 1:00if we give agents a bit more freedom.
31. 1:03For these reasons, the whole setup is
32. 1:05based on a VPS where I keep all my
33. 1:07projects.
34. 1:09Grokbot connects to the VPS securely via
35. 1:12Tailscale.
36. 1:13It’s a free tool that allows you to
37. 1:15connect devices into one private network
38. 1:17.
39. 1:17This way, my cloud computer and
40. 1:19Grokbot’s computer see each other as if
41. 1:21they were on the same home network, and
42. 1:24all communication between them is
43. 1:25encrypted.
44. 1:27Beyond the cloud setup, it was also
45. 1:29important for me not to have to babysit
46. 1:31my agents. I wanted the ability to run
47. 1:34multiple sessions simultaneously, while
48. 1:36keeping things simple and communicating
49. 1:38with only one agent. So, the system
50. 1:41should orchestrate the agents itself and
51. 1:43allow them to verify their own work.
52. 1:46If the solution the agents prepare
53. 1:48doesn't match what we initially agreed
54. 1:50upon, the agents should try to correct
55. 1:52the course and fix their own errors.
56. 1:56My role is therefore limited to setting
57. 1:58the direction for what to work on and
58. 2:00iterating on new versions of a given
59. 2:02functionality.
60. 2:04Let's see how this looks in practice. We
61. 2:06will work on a new feature for my
62. 2:08Diagram Framework application.
63. 2:10It is an application for building and
64. 2:12editing animated diagrams for my videos
65. 2:14. The whole thing was built using the
66. 2:16approach we are going through today. We
67. 2:18would like to add support for sound
68. 2:20effects to the animations. We will pull
69. 2:22the sounds from Epidemic Sound, which
70. 2:24provides an MCP for our agents.
71. 2:25Everything should be synchronized with
72. 2:28the animation tempo, appearing elements
73. 2:30, and transitions, and I should be able
74. 2:32to edit each effect.
75. 2:34Because after receiving instructions, it
76. 2:37connects to the VPS and launches Orka in
77. 2:39the terminal, my custom CLI tool for
78. 2:41managing agent work. Orka will oversee
79. 2:44the entire process, from the initial
80. 2:46idea through the agents 'work and
81. 2:47verification, all the way to reporting
82. 2:50task completion back to the bot. First,
83. 2:52Orka prepares the working environment.
84. 2:54It creates a new branch and worktree.
85. 2:57Next, it launches a new Herdr session
86. 2:59where the agents will work. Herdr allows
87. 3:02us to manage multiple terminal sessions
88. 3:04simultaneously.
89. 3:06It enables organization into projects or
90. 3:08tabs where the agents will operate.
91. 3:10It has native support for VPS.
92. 3:13This ensures that agents keep working
93. 3:15even if the connection to the server is
94. 3:17lost.
95. 3:19Thanks to it, I can also observe how
96. 3:21agents work on the VPS from my own
97. 3:23computer.
98. 3:25Statuses are another great feature.
99. 3:28Herdr monitors every agent session and
100. 3:30can tell us the status of any given
101. 3:32agent.
102. 3:33This lets us know when an agent, for
103. 3:35instance, finishes its work or gets
104. 3:37blocked by something. Orka uses the
105. 3:39Herdr CLI to launch a new agent session
106. 3:41.
107. 3:42In our case, depending on the needs, we
108. 3:45use either Claude or Spai. The main
109. 3:47agent managing the entire task is the
110. 3:49so-called orchestrator, usually an agent
111. 3:52based on a better, smarter model. It
112. 3:54receives the entire context and goal of
113. 3:56the task. It determines the plan, the
114. 3:58approach to solving the problem, and the
115. 4:00verification method.
116. 4:02It will be the only one the bot
117. 4:04communicates with during a given task.
118. 4:06In our case, the orchestrator stopped
119. 4:08planning because it noticed it would
120. 4:10need an API key for Epidemic Sound.
121. 4:13Through access to statuses, Orka caught
122. 4:15that the agent was blocked and needed
123. 4:17our input.
124. 4:19It contacted Grokbot, which processed
125. 4:21the message and displayed a window for
126. 4:23us to securely provide the key. Once we
127. 4:26did that, the orchestrator finished
128. 4:28planning and broke the task down into
129. 4:30smaller subtasks. For each one, it added
130. 4:33context, a goal, acceptance criteria,
131. 4:35scope, and verification methods. Then,
132. 4:38using Orka, the orchestrator launches
133. 4:40the so-called workers, i.e., agents that
134. 4:43will perform the created tasks. We can
135. 4:46think of workers somewhat like
136. 4:47sub-agents, but with a bit more control
137. 4:49.
138. 4:50Each one appears in a separate window
139. 4:52and receives a special prompt with
140. 4:54instructions and the task context.
141. 4:57When an agent finishes its work, Orka
142. 4:59notifies the orchestrator. It then
143. 5:01decides what to do next. For example, it
144. 5:03can send a message to a worker with some
145. 5:06corrections or spawn another worker.
146. 5:08Are you looking for a valuable place to
147. 5:10learn about AI? Is it hard for you to
148. 5:12keep up with all these changes? For six
149. 5:14years, I’ve been running a newsletter
150. 5:16where I share interesting things about
151. 5:17AI and programming. I write about what
152. 5:19works for me and how I get the most out
153. 5:21of AI. There’s one rule: zero
154. 5:24and spam. Several thousand people are
155. 5:26already with us; it costs nothing, and
156. 5:28you can always unsubscribe. Check out
157. 5:30the first link in the description, and
158. 5:31let’s get back to the video.
159. 5:33Okay, but how will we even know that the
160. 5:35workers did their job well and that the
161. 5:37feature works as we planned?
162. 5:40That’s where a crucial piece of this
163. 5:42puzzle comes in: a special agent run by
164. 5:44the orchestrator that verifies the
165. 5:46workers' output. Its goal is to click
166. 5:49through the application and gather
167. 5:50evidence that a given functionality
168. 5:52actually works.
169. 5:54Under the hood, the agent uses a library
170. 5:56to interact with the browser, for
171. 5:58instance, to record videos or take
172. 6:01screenshots.
173. 6:02Also important are the so-called feature
174. 6:04maps—markdown files that contain
175. 6:06descriptions of how our app works and
176. 6:08how the agent can perform specific
177. 6:10actions.
178. 6:11In our case, we want the agent to know
179. 6:13how the animation editor, the timeline,
180. 6:16and the playback controls work.
181. 6:19When we introduce a new feature, first,
182. 6:21we don't want to introduce any
183. 6:23regressions, but second, we must ensure
184. 6:25it integrates properly with other parts
185. 6:27of the system.
186. 6:29If the agent detects that something
187. 6:30isn't working, it notifies the
188. 6:32orchestrator, providing context and
189. 6:34evidence. The orchestrator sends a
190. 6:36message to the worker, and the worker
191. 6:38fixes the bug. After the fix, another
192. 6:39round of tests is run, and only when
193. 6:41everything works as it should is the
194. 6:43task considered finished.
195. 6:46Okay, it's cool that it works at all,
196. 6:48but won't we end up with one giant pile
197. 6:50of slop this way?
198. 6:52How can we ensure the quality of such
199. 6:54solutions and how can we increase our
200. 6:56trust in the agents 'work? From an
201. 6:58orchestration level, Orka ensures that
202. 7:01sessions do not exceed 40%of the context
203. 7:03window.
204. 7:04When that happens, a handoff to a new,
205. 7:07fresh agent occurs, and it simply
206. 7:09continues the work.
207. 7:11However, a lot depends on what our
208. 7:13current codebase looks like and how we
209. 7:15ensure agents create code the way we
210. 7:17want.
211. 7:18There are at least several ways to
212. 7:20verify this.
213. 7:22We're talking about files for agents,
214. 7:24skills, but also more deterministic
215. 7:27methods, such as linter rules.
216. 7:30On top of that, there is the review
217. 7:32process, regardless of whether we
218. 7:34perform it locally or at the pull
219. 7:35request level.
220. 7:37When the task is fully completed and
221. 7:39tested, Orka communicates with Grokbot
222. 7:41via a webhook, and it sends me a summary
223. 7:43with screenshots of the finished feature
224. 7:45.
225. 7:47This whole approach is very flexible and
226. 7:49open to modifications, so you can adapt
227. 7:51it to your own needs.
228. 7:53I often use this system by recording
229. 7:56long, casual voice notes for Grokbot
230. 7:58with entire ideas for apps or specific
231. 8:00features.
232. 8:02It runs those sessions in the background
233. 8:04and simply gets back to me when it's
234. 8:06done.
235. 8:07I also have a "night shift" system,
236. 8:08where I tell the agent I'm going to
237. 8:10sleep and I just want specific things
238. 8:12done by the morning.
239. 8:15This usually works quite well, because
240. 8:17the agent has quite a lot of time to
241. 8:19verify and process all of it. I was
242. 8:21inspired by, among others, Loren’s
243. 8:23approach, who works daily on developing
244. 8:25GregBot, and Adam Gospodarczyk’s project
245. 8:27, which works on a very similar
246. 8:29principle. I am constantly developing
247. 8:31this system myself and fixing various
248. 8:33shortcomings. You can find the link to
249. 8:35the repository in the description. If
250. 8:36you liked it, don't forget to like this
251. 8:38video and subscribe to stay up to date.
252. 8:40See you next time.