Why I switched to Pi... — Transcript

Full transcript

1. 0:00Thanks HubSpot for sponsoring this
2. 0:02video.
3. 0:03If you're into AI agents, you've
4. 0:05probably heard about this Pi agent quite
5. 0:07a lot recently. And most of you probably
6. 0:08know Pi agent because it was what
7. 0:10powering Open Claw. And today I want to
8. 0:12give you a breakdown of what's special
9. 0:14about Pi agent, why it is one of the
10. 0:16best option for you to build agent
11. 0:18system, and in the end how can you use
12. 0:19Pi to build some agentic system like
13. 0:22Post AI from scratch, which is agent
14. 0:24that designed to launching and running a
15. 0:26business. But before we dive into Pi, I
16. 0:28know a lot of you already using coding
17. 0:29agent like Cloud Code or Codec everyday.
18. 0:31But this probably a 10 or even 100X
19. 0:34difference between someone who just open
20. 0:36terminal, send a prompt, and hopes for
21. 0:38the best versus someone who fully
22. 0:40utilize every aspect of what this modern
23. 0:42harness provides. Things like what type
24. 0:44of Cloud or MD or agent.md actually
25. 0:46works, how you structure code base that
26. 0:48is agent ready, what are hooks that up,
27. 0:50and how to keep a long running work on
28. 0:51track with task state, and how to run
29. 0:54more post sessions without creating a
30. 0:55complete mess. If you want to be on the
31. 0:57right side of gap, then this free Cloud
32. 0:59Code Playbook I put together with
33. 1:01HubSpot will get you there. This is a
34. 1:03playbook I wish I had when I was
35. 1:04figuring all this out the hard way. It
36. 1:06covers my prompting patterns, how I set
37. 1:08up my Cloud or MD and project context,
38. 1:10which hooks and guardrails I use, and
39. 1:12how I manage the task state for longer
40. 1:14work. My favorite part is probably the
41. 1:16parallel sessions setup that enable you
42. 1:17to run not just one or two sessions, but
43. 1:2010 to 15 different agent sessions safely
44. 1:22at the same time. Although it is called
45. 1:23the Cloud Code Playbook, the principles
46. 1:25actually apply to Pi as well. Because
47. 1:27for every Cloud Code feature, there's
48. 1:28almost a equivalent package that you can
49. 1:31plug into your Pi agent, which we will
50. 1:33talk about soon. So most of the
51. 1:34principles here are harness agnostic,
52. 1:36which means no matter which agent you
53. 1:37use, you can adopt it. I put a link in
54. 1:39the description below so you can
55. 1:41download for free. And thanks HubSpot
56. 1:42for sponsoring this video. Now, let's
57. 1:44get back to the Pi agent. So
58. 1:45fundamentally, Pi agent is a coding
59. 1:47agent. And every time when people ask
60. 1:48me, "Hey, is Cloud Code better or Codec
61. 1:50better?" My answer for the past few
62. 1:52weeks just they are same. There's no
63. 1:54real difference between those coding
64. 1:55agents anymore. But why did Open Claw
65. 1:57choose Pi Agent as the base to build
66. 1:59upon? Because CloudCode, CodeX,
67. 2:00OpenCode, they all have their own SDK
68. 2:02and CLI that allow you building new
69. 2:04agent system by wrapping it. Well, the
70. 2:06core difference is all the other coding
71. 2:07agents harness, even though it's really
72. 2:09powerful, you can't really modify or
73. 2:11change how the harness inside it works.
74. 2:14Even CloudCode and CodeX is becoming
75. 2:16increasingly similar to each other, but
76. 2:18each one still have their own unique
77. 2:19features. Like CloudCode has dynamic
78. 2:21workflow, but if you're user of CodeX,
79. 2:23you can't just tell CodeX to say, "Hey,
80. 2:25implement this dynamic workflow as part
81. 2:27of your harness." And vice versa. Even
82. 2:29though they already tried to open up
83. 2:31more customizability by exposing hook
84. 2:33system to the user. And if you don't
85. 2:35know hook, they normally allow you to
86. 2:36define some programmatic behavior when
87. 2:38certain things happen, like when a
88. 2:40session start or before the agent try to
89. 2:42call a tool. So, you can achieve some
90. 2:43sort of harness modifications by doing
91. 2:45things like every time when CloudCode
92. 2:47try to call a tool, it automatically
93. 2:49check a permission rules that you write
94. 2:51and then return back true or false based
95. 2:52on permission rules that you want. It's
96. 2:54pretty powerful, but still a bit
97. 2:55limited. Like, if you look at the GitHub
98. 2:57issues, there are a lot more scenarios
99. 2:59of the hooks that users want to modify,
100. 3:02but couldn't. And this is where Pi Agent
101. 3:05really shine. So, the whole design
102. 3:06philosophy of Pi Agent is that harness
103. 3:08should actually adapt to the user rather
104. 3:10than the other way around. So, at
105. 3:12default, Pi comes with a minimum version
106. 3:14of coding agent. It only has four
107. 3:16critical tools: run bash command, write,
108. 3:18read, and edit files. There's no sub
109. 3:20agents, no agent teams, and don't even
110. 3:23have MCP comes with it. It's very bare
111. 3:25minimum. But, meanwhile, it can be
112. 3:27easily extended with any other features
113. 3:29that you want. They have this concept
114. 3:30called extension. So, users and agents
115. 3:33can just write new extension file to
116. 3:35extend the Pi coding harness. They
117. 3:37expose a whole bunch of things to allow
118. 3:38you to make modifications across the
119. 3:40tools, contexts, hooks, session
120. 3:42management, command UI, almost
121. 3:44everything that you can think of. So,
122. 3:45you can add in new tools, command
123. 3:46shortcut providers, large language model
124. 3:48providers, and they They almost every
125. 3:51single hooks that you will ever need.
126. 3:53So, you can write programmatic behavior
127. 3:54to change and improve the harness. They
128. 3:56even expose the UI itself as well. For
129. 3:58example, I can just customize my Pi
130. 4:00Agent UI to whatever I want, like this
131. 4:02random character. And the cool thing is
132. 4:04that this extension can be written by
133. 4:06you as human or by the agent itself. Pi
134. 4:09Agent has its knowledge baking already
135. 4:11and has full awareness of what kind of
136. 4:13extension already loaded and how to
137. 4:15write the extension. You can just modify
138. 4:17itself on the fly. For example, I can
139. 4:19just tell it, "Hey, I want to see the
140. 4:20weather info in the prompt input. Help
141. 4:22me customize your UI." It will
142. 4:24automatically read the extension doc,
143. 4:26then write a new extension for the
144. 4:27weather widget. Once it's done, you can
145. 4:29just do {slash} reload. It will
146. 4:31immediately reload the latest version of
147. 4:33the harness that it just created with
148. 4:34those extensions. And now you can see
149. 4:36the weather information is showing up
150. 4:37here, just above my prompt input. And
151. 4:40this type of UI customization is just
152. 4:41really small part of all the
153. 4:43possibilities. You can almost ask it to
154. 4:45update things on the fly for any part of
155. 4:47harness. And there's a Pi Agent package
156. 4:49catalog. So, you can install extensions
157. 4:51built by others. So, even though default
158. 4:53Pi Agent don't have any of those
159. 4:54features, we can find almost every
160. 4:56popular Cloud Code Codex feature already
161. 4:59implemented by someone that you can plug
162. 5:01into your Pi Agent, like the Go feature,
163. 5:03the MCP adapter, the Chrome browser
164. 5:05access, the ask user question tool, sub
165. 5:07agents, plan mode, and even dynamic
166. 5:09workflow and computer use. If you want
167. 5:11to enable dynamic workflow, you simply
168. 5:13copy this package of dynamic workflow,
169. 5:15run this command, then reload. Now I can
170. 5:17just give a prompt including keywords
171. 5:19like workflow, then works almost like
172. 5:21Cloud Code's dynamic workflow feature.
173. 5:23It will trigger workflow, and I can also
174. 5:25do {slash} workflow, which will pop up
175. 5:27this workflow UI, and you just get a
176. 5:28very similar, almost identical
177. 5:30experience from the Cloud Code. There
178. 5:32even package that will just turn your Pi
179. 5:34Agent into a Cloud Code or Codex. The
180. 5:37package all special tooling those other
181. 5:39harness has, so you can install and use
182. 5:41that artifact. There even extension that
183. 5:43allow you to play DOM while the agent is
184. 5:45running. And some of the extensions
185. 5:46might not be well-built, but technically
186. 5:48you can actually ask agent update and
187. 5:50build for you. And with this
188. 5:51flexibility, I can already see some
189. 5:53really interesting things that only
190. 5:54possible to be built with pipe. For
191. 5:56example, this one package called Pi
192. 5:58hyper. What it does is that it will
193. 5:59capture every tool call hook for things
194. 6:01like bash command to actually making
195. 6:03sure it pre-processing the bash command
196. 6:06result and only return relevant and
197. 6:08minimum information back to the agent.
198. 6:10For example, previously, if agent is
199. 6:11running Git log to see what are recent
200. 6:13commits, command will include a lot of
201. 6:15information for each commit like hash,
202. 6:16author, date, body, and diff blocks. But
203. 6:19with this package, it will actually
204. 6:20clean things up to show only what
205. 6:22matter. And just by doing this, it can
206. 6:24probably reduce 96% of those type of
207. 6:26commits. And in some other testing
208. 6:28commits, it will all cut token by like
209. 6:3080 to even 90%. And this is a good
210. 6:32example where things like that is just
211. 6:33difficult to implement with cloud code
212. 6:35or codex because their pre-tool use cook
213. 6:38only allow you to append new information
214. 6:40to the tool call rather than directly
215. 6:42modify the tool call result. So, this is
216. 6:44rich extension system is a core that
217. 6:46making Pi agent a super useful and this
218. 6:48exactly how Open Claw started. So, first
219. 6:51what they were building is mainly the
220. 6:52gateway to connect with app Slack and
221. 6:54the web interface while using the Pi
222. 6:56agent as the main agent runtime. So, you
223. 6:58can utilize agent loops, large language
224. 7:00model OS, and the session management,
225. 7:02but also extend it with MCP, memory sub
226. 7:04agent, ACP, and many other things. And
227. 7:06those type of customizability would be
228. 7:08otherwise very difficult to achieve if
229. 7:10Open Claw were using cloud code SDK or
230. 7:12codex CLI. And it's actually pretty easy
231. 7:14to build those extensions. I'm going to
232. 7:16show you a few quick examples. So, to
233. 7:18build extension, you can just create a
234. 7:19dot pi folder in any project folder or
235. 7:21in your root and then add an extensions
236. 7:23folder inside it. And here is a few
237. 7:25examples. So, let's say you want to
238. 7:26modify agent's context to always be
239. 7:28aware of the current Git repo like which
240. 7:31branch it is on, what are the available
241. 7:32work trees, un-staged changes, and
242. 7:34recent commits. We can simply create
243. 7:36extension like this, which is in
244. 7:37TypeScript, and adding a Git function
245. 7:39which will run the Git command get all
246. 7:41those Git related information back,
247. 7:43extract and generate a Git summary. And
248. 7:45then just simply do this
249. 7:46pi.on_before_agent_start
250. 7:48and modify the system prompt, append
251. 7:50additional Git context after the
252. 7:52predefined system prompt. And this is as
253. 7:54simple as it get. Once you put this
254. 7:56extension file here, the next time you
255. 7:58load the pi agent, it will automatically
256. 7:59have those contexts. I can simply ask
257. 8:01it, "What's the Git information here?
258. 8:03Use no tools to answer me." And it will
259. 8:05return back the information. And
260. 8:06meanwhile, you can also easily add
261. 8:07custom tools. Like, what if you want
262. 8:09your pi agent to be able to read what's
263. 8:11in your clipboard? You can simply do
264. 8:13this pi.register_tool,
265. 8:15give it a name, label, description, and
266. 8:17param, then define the actual function
267. 8:19here. With this, if I go copy the
268. 8:21previous Git context extension file, and
269. 8:23I will just ask it, "What's in my
270. 8:24clipboard?" It will use the tool that we
271. 8:26write to get clipboard information. And
272. 8:28we can even build something more
273. 8:29interesting, like this permission gate.
274. 8:31So, the idea is that what if you want
275. 8:32this agent to be used by your team,
276. 8:34where each one might have different
277. 8:35permission access to different level of
278. 8:37information? Instead of message sent
279. 8:39directly to agent, or let the agent
280. 8:41decide whether the user has right
281. 8:43permission, what if we add a permission
282. 8:45check using a cheap and fast small
283. 8:46model? If it's all good, then continue
284. 8:49with agent as normal. But if it is some
285. 8:51dangerous information, you can just stop
286. 8:52the session right away. And here, we
287. 8:54just write another extension. We are
288. 8:55using two different package, which I
289. 8:57will dive a bit deeper. One is a pi
290. 8:59coding agent that allow you to write
291. 9:00extension API, but also this pi AI
292. 9:04package, which is some package from pi
293. 9:05agent that allow you to do large
294. 9:07language model call with any model very
295. 9:08easily. So, we define a small gate model
296. 9:11to use Haiku and output decision whether
297. 9:13it's allowed or denied, plus a
298. 9:15reasoning. Then we package this as small
299. 9:16decide tool, the system prompt, as well
300. 9:19as the user message, which load up both
301. 9:21the policy, which is permission.md file
302. 9:23that user can define on the root, plus
303. 9:25the user request. And we use this
304. 9:26function provided by the pi AI package
305. 9:29to direct call the large language model
306. 9:31and output result. All we need to do is
307. 9:33just when the session start, we will
308. 9:35load up the policy information in the
309. 9:37prompt. And then after every time user
310. 9:39send a message, we're showing message on
311. 9:41the UI saying permission checking, wait
312. 9:43for the large language model result. If
313. 9:45it is denied, then on the UI we will
314. 9:47show this block information and skips
315. 9:49agent entirely. Otherwise, continue like
316. 9:51normal. To test it, you can just define
317. 9:53a permission.md file to define a
318. 9:54specific permission. Like here I said
319. 9:56only Jason can get revenue data. So if I
320. 9:59go to my Pi agent, ask what's the
321. 10:01revenue for June, it will trigger
322. 10:02permission and show me back this
323. 10:04information that is blocked. That you
324. 10:06don't have permission to access revenue
325. 10:07information. So this is one example of
326. 10:09how you can update Pi agent's harness by
327. 10:12writing those extension. And most of the
328. 10:13time you actually don't need to learn
329. 10:15building those extension yourself, cuz
330. 10:17you can just talk to the Pi agent, it
331. 10:18will self-evolve. However, the real
332. 10:20power of Pi, from my point of view, is
333. 10:23actually not just use it as a coding
334. 10:25agent, but actually build on top of Pi
335. 10:27agent to powering your AI products. If
336. 10:29you can look at Pi agent repo, it
337. 10:31actually comes with five different
338. 10:32package. The first one is this AI
339. 10:34package. This is almost like a Vercel AI
340. 10:37SDK that allow you to call different
341. 10:39large language models. It comes with all
342. 10:41the OAuth function to allow your user to
343. 10:43connect cloud code subscription, code
344. 10:45act subscription. So you can actually
345. 10:46use this as standalone function to just
346. 10:49do large language model call in the
347. 10:50application you're building. The second
348. 10:51package is this agent package. This is
349. 10:54basically the agent loop. You can
350. 10:55consider almost this like a Vercel AI
351. 10:57SDK stream text function that has this
352. 11:00small agent loop built in. The third one
353. 11:02is this coding agent. So coding agent
354. 11:04you can use agent core loop as well as
355. 11:07AI package, but comes with this
356. 11:09essential tooling like read, write, add,
357. 11:11and bash, as well as session management,
358. 11:13compassionate context window, the
359. 11:15extension system, as well as SDK. And
360. 11:17this package, you can almost consider as
361. 11:20equivalent to the cloud code agent SDK,
362. 11:23but it is fully customizable and you can
363. 11:25connect to any model you want. And in
364. 11:26the end it also comes with this TUI
365. 11:28package that wraps the coding agent in
366. 11:30this terminal UI. And also a up trader
367. 11:33package. So this up trader package allow
368. 11:34you to build like schedule job or
369. 11:36delegate task to different Pi agent
370. 11:38process. And this is been more
371. 11:39experimental. But you can see that based
372. 11:42on the type of a product you're
373. 11:43building, you can actually utilize the
374. 11:44Pi agent's ecosystem to scaffolding the
375. 11:47agent products. And I actually utilize
376. 11:49this coding agent SDK to build a product
377. 11:52like Posia, which is a a times agent
378. 11:54that can launch and run business to
379. 11:55showcase the real power. And I'm going
380. 11:57to show you very quickly how I build it
381. 11:59to help you draw a picture. So as I
382. 12:01mentioned, I'm using this coding agent
383. 12:02package because it comes with a SDK that
384. 12:05you can use. If you are trying to build
385. 12:07a local agent, which means the agent
386. 12:09system that lives on user's own
387. 12:11computer, then this package basically
388. 12:12give you most of stuff that you need.
389. 12:14And you can just easily extend
390. 12:15functionality by building extension
391. 12:17system, custom UI, and task triggers.
392. 12:20One good example is this all my Pi
393. 12:22coding agent. This is coding agent that
394. 12:24build on top of Pi package. But it comes
395. 12:26with a lot of new tools that brings
396. 12:28performance closer to the cloud code
397. 12:30Codex. It comes with editor type of user
398. 12:33interface inside terminal. It even have
399. 12:35features that allow you to just publish
400. 12:37a specific conversation session you have
401. 12:39with someone else. It has a QR code that
402. 12:41people can scan or a URL you can share
403. 12:43with others. So you can actually build
404. 12:45very sophisticated local agent products
405. 12:47on top of Pi agent. But meanwhile, you
406. 12:49can also use Pi agent package to build
407. 12:52web hosted agent product like this Posia
408. 12:54replica I did. But there are some nuance
409. 12:56you need to handle because the Pi
410. 12:58e-coding agent SDK at default is
411. 13:01designed quite tied to the local file
412. 13:02system. But when you have this web
413. 13:04deployed agent, which means agent run
414. 13:06actually lives on your shared backing
415. 13:07infrastructure. But when you try to
416. 13:09write file, run bash command, then each
417. 13:11user has their own sandbox. With this
418. 13:13type structure, you will need to make
419. 13:14some modifications. Like instead of
420. 13:16using Pi coding agent default session
421. 13:18manager system, which is directly tied
422. 13:20to the general file on the computer, you
423. 13:22probably want to store the session
424. 13:24information in your DB and manage
425. 13:25compassion yourself. And also for the
426. 13:27default bash added read write tools, you
427. 13:29probably want to wrap a layer so that it
428. 13:31will do those operations in the sandbox
429. 13:33that user is owning. But, it is same
430. 13:35situation we need to handle even though
431. 13:37you are using cloud code agent SDK. And
432. 13:39with this, you can also build a web
433. 13:41application like this, which is a chat
434. 13:43agent. User can chat to it. To implement
435. 13:45such experience, we pretty much still
436. 13:47using the pi coding agent package. But
437. 13:49this time we use the default resource
438. 13:50loader. So, this is how you can pass
439. 13:52additional extension skill to the coding
440. 13:54agent. For example, here we define a
441. 13:56extension for guardrail that it will
442. 13:58block by running certain bash commands.
443. 14:01And also define a tool to get MRR with
444. 14:03some mock result. Then we define this
445. 14:05resource loader. So, we can pass on the
446. 14:08directory the agent should be working in
447. 14:10and in the directory where agent should
448. 14:11load extension file like OS key. And you
449. 14:14can also define whether the agent should
450. 14:16just load up extension skills from the
451. 14:18door pi folder. We can also passing
452. 14:20customized extensions. Like here, we are
453. 14:22passing the guardrail's extension that
454. 14:24we define in line here. As well as
455. 14:26additional skills. So, this will allow
456. 14:27you to customize what type of skills
457. 14:29each agent session should have access
458. 14:31to. Then use this create agent session
459. 14:33function to passing the custom tools as
460. 14:35well as a resource loader. And the rest
461. 14:37will be similar. Based on the type of
462. 14:39events, you can present different
463. 14:40information on the UI. So, this is how
464. 14:42you can build a wrong pi package. I
465. 14:44actually use the same pi SDK build a
466. 14:46replica version of post here, which is
467. 14:47hosted agent system that can launch and
468. 14:49run company autonomously. It is using
469. 14:51pi's coding agent SDK as agent runtime.
470. 14:54Breaking down into 11 different agents
471. 14:56with orchestrator that's built around
472. 14:58task entity and persist the state and
473. 14:59context as well as tools proxy. I'll
474. 15:01share more details step-by-step tutorial
475. 15:03of how did I build the system from
476. 15:05scratch in a builder club. So, you can
477. 15:06join if you're interested. Meanwhile,
478. 15:08I'll add some of the extension I build
479. 15:09or some of my own favorite pi extensions
480. 15:12into our curious skill folder. So, you
481. 15:14can give it to pi agent and have it set
482. 15:16up. I hope this is useful. Thank you.
483. 15:17And I'll see you next time.