What is an AI harness? I build one live in less than 30 minutes — Transcript

Full transcript

1. 0:00A harness is some code around an AI
2. 0:03agent that makes it more effective. Why
3. 0:06we've seen people build these specific
4. 0:08use case harnesses is sometimes with the
5. 0:11specific job, you just want to
6. 0:14micromanage a little bit. You just want
7. 0:15to be more prescriptive about how that
8. 0:18job gets done. [music] I'm going to show
9. 0:19you how it works and then we will talk
10. 0:22about how I built it. So the interface I
11. 0:25built for my harness is a terminal UI.
12. 0:28The harness core is run on cloud
13. 0:30agent SDK and then it's connected to
14. 0:32real tools. So it's connected to [music]
15. 0:34Sentry, Vercel, and then it's connected
16. 0:36to linear and GitHub in terms of getting
17. 0:38[music] tasks done. I think we all have
18. 0:40done good work, but then now I've
19. 0:42realized that these agents can help us
20. 0:44solve very very specific problems by
21. 0:46constraining that work. It's really like
22. 0:48changed my mind about how work gets
23. 0:51done.
24. 0:52>> [music]
25. 0:54>> Everybody's saying it's not the model,
26. 0:56it's the harness. But you know what not
27. 0:59everybody is saying?
28. 1:00What is [music] a harness? In today's
29. 1:03how I AI episode, I'm going to demystify
30. 1:07the idea of a harness, write my own
31. 1:09harness and show you how you can do the
32. 1:11same, and explain to you why a custom
33. 1:14harness makes sense and could be better
34. 1:16than using Claude code or Codex alone.
35. 1:19Let's get to it. This episode is brought
36. 1:21to you by Bolt.new, the AI app builder
37. 1:24for people who have ideas and want to
38. 1:26ship them.
39. 1:27Most AI tools spit out code that looks
40. 1:30great in the demo and falls apart the
41. 1:31second [music] you try to do anything
42. 1:33real with it. Or they lock you into
43. 1:36their own platform with no real way out.
44. 1:38Bolt is different. [music] You describe
45. 1:40what you want to build, a startup MVP, a
46. 1:42landing page, an internal tool, a side
47. 1:44project, and Bolt generates
48. 1:46production-ready code in minutes.
49. 1:49Connect Stripe or any other MCP, hook up
50. 1:51your domain, and deploy it live.
51. 1:54Founders are using Bolt to build
52. 1:55businesses doing real [music] revenue.
53. 1:58Product managers are shipping prototypes
54. 2:00their teams actually use. Designers and
55. 2:03marketers are launching campaigns
56. 2:05without waiting in line. Anyone can
57. 2:07build, engineering can ship, everyone
58. 2:10[music] wins. You just need an idea and
59. 2:12a weekend. Check it out at
60. 2:14bolt.new/howiai.
61. 2:18Before I get into how to build a
62. 2:20harness, let's talk about what a harness
63. 2:22is and I'm going to make it as simple as
64. 2:26I can for all of you. A harness is some
65. 2:29code around an AI agent. Yes, you heard
66. 2:32it here first. A harness is just code
67. 2:36around an AI agent that makes it more
68. 2:38effective. Can that code have AI in it?
69. 2:41Sure. Does that code have to have AI in
70. 2:44it? Not necessarily. What is the goal of
71. 2:47a harness? To make the AI better. It is
72. 2:49so simple and I feel like the way that
73. 2:52people have been talking about this have
74. 2:54made it such a mystery that I wanted to
75. 2:56make it just very clear to you all. It
76. 2:59is just writing more code around your AI
77. 3:01to make it more useful for a specific
78. 3:04use case.
79. 3:05So, what are the parts of a harness?
80. 3:07Well, a harness is going to have
81. 3:09specific context, it's going to be able
82. 3:11to take specific actions, and it's going
83. 3:13to have a goal of specific outcomes.
84. 3:16It's just as simple as that. And I want
85. 3:20to talk about when it makes sense to
86. 3:22build a harness and when it doesn't. And
87. 3:26I think you want to build a harness when
88. 3:28the same workflow needs the same setup
89. 3:32and the same outcomes. And so, it's kind
90. 3:35of similar to when you would build an AI
91. 3:37agent. In fact, harness agent, sometimes
92. 3:39you can interchange some of these
93. 3:41concepts, but really it's when there is
94. 3:44a sort of combination of deterministic
95. 3:47and non-deterministic workflow,
96. 3:48step-by-step process tools use cases,
97. 3:52you want your AI to follow up to do a
98. 3:55specific job. Usually those jobs are
99. 3:57like slightly more complex. And this is
100. 4:00why you've seen these coding harnesses
101. 4:02come out like coding is a job to be
102. 4:04done. It needs specific tools. It
103. 4:06typically goes through kind of a
104. 4:07standard workflow. And so coding
105. 4:09harnesses are very popular, but you
106. 4:12could also do things like managing
107. 4:14production incidents where you need to
108. 4:15go through a specific process, getting
109. 4:17PRs ready for release, of handling
110. 4:20support escalations, managing
111. 4:22migrations.
112. 4:24Even non-technical use cases like doing
113. 4:26research in a very specific way or
114. 4:28consolidating docs in a very specific
115. 4:30way. That's how you and why you would
116. 4:31use a harness.
117. 4:34So, how did I decide what kind of
118. 4:37harness I would build? Well, I looked
119. 4:39across my business at Chat Priority and
120. 4:41I thought, "What am I doing sort of
121. 4:44repeatedly and consistently that I think
122. 4:46AI could be good at? That I think we
123. 4:48could be doing better if we were more
124. 4:50structured about the AI and how we used
125. 4:52it." And I thought that fixing bugs, you
126. 4:57all if you've listened to this podcast,
127. 4:58look, I ship code so I ship bugs. Fixing
128. 5:00bugs is a very specific workflow where
129. 5:04we've built some custom internal tools
130. 5:06that I've been generally doing with
131. 5:07Claude Coder Codex, but I had a had this
132. 5:09hypothesis that I could do a better job
133. 5:12of triaging bugs if I built my own
134. 5:14harness. And so, I picked Sentry
135. 5:18debugging and sorry for the Claude slot
136. 5:21content here. Um
137. 5:23Sentry debugging and debugging Sentry
138. 5:25issues, really figuring out the issue,
139. 5:27using some of our custom internal tools,
140. 5:30and then doing all the follow-up actions
141. 5:33we do when we close bugs was like a good
142. 5:35first harness. It had coding in it. It
143. 5:38needed custom content and custom
144. 5:40context. There were like specific
145. 5:42outcomes I wanted to make sure that we
146. 5:44followed like tracking everything in
147. 5:46linear and writing follow-up docs that
148. 5:48the rest of the engineering team could
149. 5:49use. And so, we chose uh debugging our
150. 5:53Sentry bugs, by we I mean me and Codex,
151. 5:55chose debugging Sentry as a good use
152. 5:58case to demonstrate how to build a
153. 6:00harness.
154. 6:01Now, why wouldn't I just use an AI
155. 6:03coding tool directly? Well, I have been
156. 6:05using AI coding tools directly. And I
157. 6:08think the problem with using a
158. 6:09general-purpose coding tool and why
159. 6:12we've seen people build these specific
160. 6:14use case harnesses is sometimes with a
161. 6:17specific job, you just want to
162. 6:20micromanage a little bit. You just want
163. 6:21to be more prescriptive about how that
164. 6:24job gets done. And so, if you can
165. 6:26identify the right workflows, you can
166. 6:28actually be more efficient, more
167. 6:30consistent, and have better outcomes if
168. 6:34you build a harness. So, for this
169. 6:36specific use case, you know, with a
170. 6:39direct AI tool like Claude Code, um I
171. 6:42would have to explain what I want the
172. 6:45the agent to do. So, I'd say like, "Dear
173. 6:47agent, please fix this bug. Here it is."
174. 6:50and send a link. Instead of this
175. 6:51harness, I can literally just paste in
176. 6:53the link and the agent already knows my
177. 6:55intent, already knows what the job to be
178. 6:57done.
179. 6:58A second thing that I wasn't that
180. 6:59worried about, but is interesting when
181. 7:01you build your harnesses, you can be
182. 7:02really prescriptive about what tools
183. 7:04it's allowed to do and what it's allowed
184. 7:05to execute and not. So, for example, if
185. 7:08you wanted to build an investigate-only
186. 7:11harness, you could make sure that your
187. 7:15harness, your code editor, never
188. 7:17actually wrote code. It only explored
189. 7:20and explained root cause. You can also
190. 7:22repeat the same process over time if you
191. 7:25encode it in a harness. And so, if you
192. 7:27want like a very precise step-by-step
193. 7:30flow, including outcomes, so for us,
194. 7:32every time we fixed a
195. 7:35Sentry bug, we want it documented in
196. 7:36Linear, we want a very specific report,
197. 7:39we might even want to follow up with
198. 7:40customers that it was impacted with.
199. 7:42You could encode that in a skill, but
200. 7:44then again, you have to babysit it. When
201. 7:46we build this harness, we knew it would
202. 7:47happen every time. And then, from a
203. 7:51model perspective, you can do multimodal
204. 7:53routing and all sorts of interesting
205. 7:54things and ways that you couldn't with a
206. 7:56general-purpose AI model. So, I'm going
207. 7:59to show you how it works, and then we
208. 8:01will talk about how I built it. Okay, so
209. 8:05the interface I built for my harness is
210. 8:07a terminal UI, again, like Claude Code
211. 8:09or Code Geeks, something you would run
212. 8:10in your eye in a UI.
213. 8:13And just so you know,
214. 8:14your harness does not have to be a TUI.
215. 8:17It doesn't have to be a CLI. It doesn't
216. 8:19even have to have letters. It could be a
217. 8:21web app.
218. 8:22I did it in a TUI one because I haven't
219. 8:24built one in a while. I thought it would
220. 8:25be fun. And two, I just want to show
221. 8:27that building your own custom harness
222. 8:29means you can build your own custom
223. 8:31interface into these AI agents as well.
224. 8:34So, the harness is the whole experience,
225. 8:37including the human experience that
226. 8:39makes it more useful and easier to use.
227. 8:42And so, um this TUI is pretty easy to
228. 8:45invoke. I just run TUI. You can see it
229. 8:48here. It's kind of cute. It's been made
230. 8:50cute. Um I use this library called Ink,
231. 8:53which helps you make cute TUIs. I don't
232. 8:55think they would say cute, but I'm going
233. 8:56to say cute. And you can see here that
234. 8:59this terminal UI really reflects the
235. 9:01structure of the harness itself. So, you
236. 9:03see all the runs um that it's done so
237. 9:06far,
238. 9:08errors and how it's fixed things. And
239. 9:10then, sort of our process, which is it
240. 9:12gathers evidence, it streams in
241. 9:14activities, and then it builds some
242. 9:16artifacts. And so, I'm going to actually
243. 9:18have it investigate this Sentry error
244. 9:20over here. It's one where our edit um
245. 9:24operations are getting dropped sometime
246. 9:26by the agents. And that has now kicked
247. 9:30off our specific harness. So, what it's
248. 9:33going to do is it's going to start this
249. 9:34investigation
250. 9:36run. It's going to kick off a Claude SDK
251. 9:38session, which is a fundamental part of
252. 9:40how I built this. It's going to go ahead
253. 9:42and start gathering evidence and coming
254. 9:44up with a root cause hypothesis of
255. 9:47what's causing this issue and how we
256. 9:49might fix it. Now, as you can see, I
257. 9:51chose I investigate, not F fix. So, the
258. 9:55investigation should not touch and
259. 9:58modify files. And again, this is
260. 10:00something that I would have had to like
261. 10:02prompt to the agent and say, I only want
262. 10:04you to investigate. I do not want you to
263. 10:06ship a fix. But instead, I can just
264. 10:08click I, paste in that Sentry issue, and
265. 10:12it's off to the races. This episode is
266. 10:16brought to you by Customer.io.
267. 10:18You're here because you'd rather use AI
268. 10:20than talk about it. With Customer.io,
269. 10:23you describe the campaign you want to
270. 10:25build, and the AI agent creates for you
271. 10:28the audience, the messages, and the
272. 10:30timing. You review it, make any changes
273. 10:33you want, and launch.
274. 10:35Instead of spending hours stitching
275. 10:36together tools and workflows, you can
276. 10:38focus on the work that actually drives
277. 10:40growth. Every campaign is tied back to
278. 10:43results, so you can see what's working
279. 10:45and what to do next. [music]
280. 10:47More than 9,000 brands use Customer.io
281. 10:50to turn the data they already have into
282. 10:52messages customers remember.
283. 10:54Visit customer.io/howiai
284. 10:59to try it today.
285. 11:01Customer.io, more impact from every
286. 11:04message. While this is running, I'm
287. 11:06going to just go and show you a little
288. 11:08bit about how this works and how I have
289. 11:11actually built it. Okay, so this is the
290. 11:13high-level architecture of the app. So,
291. 11:16the front end is a terminal UI or a C
292. 11:19CLI. Each invocation of the harness we
293. 11:23call a run, so it's running a task. Each
294. 11:26task has a specific input, usually
295. 11:28that's a Sentry issue. And then there
296. 11:31are specific flags I put on the harness
297. 11:33that allow it to edit the source, modify
298. 11:38the inputs, or even message customers
299. 11:40only if I flag and approve it. So again,
300. 11:43this is just a little bit more control
301. 11:44over how the agent works. The harness
302. 11:47core is run on Claude agent SDK and so
303. 11:51all the agentic planning is run through
304. 11:53the Claude agent SDK which has some of
305. 11:56the primitives of Claude code including
306. 11:58wrapping files and writing files and all
307. 12:00those sorts of things that we find
308. 12:01useful.
309. 12:02And then what's really interesting about
310. 12:04this harness, and you've seen in other
311. 12:06harnesses like open claw, is it can
312. 12:09create its own artifacts in its file
313. 12:12store. And so we have this artifact
314. 12:14store, I will show it to you in a
315. 12:15minute, and it basically saves all the
316. 12:18evidence from these runs to the file
317. 12:20system for the agent to use in the
318. 12:23future. And then it's connected to real
319. 12:24tools, so it's connected to century, for
320. 12:26sale, the Claude SDK, it's running
321. 12:29Sonnet 46. I think that's the right the
322. 12:31right model for the job. And then it's
323. 12:32connected to linear and GitHub in terms
324. 12:34of getting tasks done. Now what's really
325. 12:38interesting as well is you can prompt
326. 12:42this in a custom way. So instead of the
327. 12:44general like you are Claude code, make
328. 12:46no mistakes, you are our, you know, sort
329. 12:49of model genius.
330. 12:51I'm saying specifically that you're
331. 12:53working inside the Chat Purity
332. 12:54engineering harness, it's Chat Purity
333. 12:56specific, it's not an open-ended coding
334. 12:58system. We want to use these artifacts
335. 13:02as a source of truth and here's the plan
336. 13:04to attack a very specific problem. And
337. 13:08what I want you to return is X, Y, and
338. 13:10Z. And again, I don't have to copy and
339. 13:12paste this, I don't even have to put it
340. 13:13in a skill where hopefully it will get
341. 13:15invoked in the right way. I've actually
342. 13:17encoded this in a very specific step in
343. 13:20the harness to make sure that the model
344. 13:22falls it every time. And so
345. 13:25there's several of these types of custom
346. 13:27prompts inside my harness. There is um
347. 13:30the artifacts that get generated. There
348. 13:32are tool policies around like what tools
349. 13:34can be called and which ones can't. And
350. 13:36then um I have just decided again to use
351. 13:39Claude Sonnet 4 6, um which is really I
352. 13:42think the right model for this
353. 13:43particular workflow.
354. 13:45Okay, I want to talk a little bit about
355. 13:47the code and how you generate this. And
356. 13:49then just like a peek behind the scenes.
357. 13:51I actually ran dueling Claude code and
358. 13:54Codex sessions and essentially said like
359. 13:57help me build a harness. I think I want
360. 13:59to use the Claude agent SDK. Here's what
361. 14:01I would like it to do. And then like
362. 14:03closed my eyes and tried to get it done.
363. 14:06Honestly, it was not a one-shot. I don't
364. 14:08know if it was my prompting or the
365. 14:11models were being funky. It was GPT 5.5
366. 14:14and Opus, but both of them really wanted
367. 14:16to build something super deterministic.
368. 14:19So, they like really resisted putting
369. 14:21any AI in the harness and I had had to
370. 14:25really prompt it very, very specifically
371. 14:28to get what I want. So, I would say if
372. 14:30you were trying to do this, I would be
373. 14:32very specific about the workflow. I
374. 14:35would be very specific about the tools.
375. 14:37I would be very specific about where
376. 14:39custom prompts make sense. And then I
377. 14:42would suggest using an agent SDK either
378. 14:45from Claude or from OpenAI to run most
379. 14:49of it because without that prompting, I
380. 14:52just did not get what I wanted out of
381. 14:54these models. The second thing I will
382. 14:55say, funnily enough, Codex did the best
383. 14:58job at building the agent, but it used
384. 15:01Claude agent's SDK to actually implement
385. 15:04the agent. So, we are spanning across
386. 15:05models and spanning across coding agents
387. 15:08here.
388. 15:09But the actual harness itself is pretty
389. 15:11simple. It's got sort of a high-level
390. 15:15index of how you get to the TUI. And
391. 15:17then it's got like, I don't know, eight
392. 15:19files of specific things it can do. So,
393. 15:23it can hunt for bugs in Sentry. Um it
394. 15:27has a Sentry adapter to effectively use
395. 15:29the Sentry API in a very specific way.
396. 15:32So, instead of using the MCP generally,
397. 15:35instead of like having your coding agent
398. 15:37wander through all these traces, I'm
399. 15:39just very precise about exactly what I
400. 15:41think you need to pull from a bug report
401. 15:42perspective, what's useful, what's not,
402. 15:44and made that connector really
403. 15:46opinionated. It's got similar a linear
404. 15:49integration and a Vercel integration and
405. 15:51a GitHub integration. So, again, not
406. 15:53like generally how you can use these
407. 15:55tools, but specifically how you would
408. 15:58use these tools when you are searching
409. 15:59for a bug. And then, after those tools
410. 16:03and data sources are used, the bug is
411. 16:05identified and triaged, then there is
412. 16:08this artifact file here that outputs and
413. 16:12spits out the specific artifact I want
414. 16:14to see after a bug run is done.
415. 16:17And that artifact bundle looks something
416. 16:19like this. So, it's literally just uh
417. 16:23the task run, um which is all the
418. 16:25messages, the report, so what was the
419. 16:28Sentry issue, here's a brief on what we
420. 16:30discovered, here's any logs that we
421. 16:33think are relevant, what the Cloud
422. 16:35Worker ended up doing, and then the
423. 16:37summary of the output. And then, we also
424. 16:40output this beautiful HTML file um that
425. 16:42I can show you that shows you
426. 16:45what happened and how it all worked, as
427. 16:47well as a worker report. So, I will show
428. 16:49you those outcomes, as well. Just
429. 16:51pulling up this code for you again,
430. 16:56it's pretty straightforward. It's giving
431. 16:58me all the instructions on where to put
432. 17:00my specific API keys, and then, I can
433. 17:03just run it in this very opinionated
434. 17:06way. So, in addition to running the TUI,
435. 17:09which lets me sort of like navigate
436. 17:10through the UI and use this harness,
437. 17:12something I might want to do as a human,
438. 17:14it also has built these really easy
439. 17:16command line tools, where if I just
440. 17:18quickly want to run this harness against
441. 17:21specific issues with specific flags on
442. 17:24tool use, I can definitely do that. And
443. 17:26what's kind of interesting about this is
444. 17:28yes, I built this harness and you can
445. 17:30see here I built this like fun UI so
446. 17:33that I could use it in a fun way and it
447. 17:35makes for a better demo, but really this
448. 17:38harness is a structured way to give
449. 17:41agents the job of running these
450. 17:44investigations on an on a simpler basis.
451. 17:46And so you can imagine while I design
452. 17:48the TUI for human, actually giving a
453. 17:52kind of
454. 17:53all intelligent agent a specific harness
455. 17:57to solve a specific problem with agents
456. 17:59in that,
457. 18:00I think that's how you're going to get
458. 18:01real leverage and really custom outcomes
459. 18:04out of things like coding agents like
460. 18:07Claude Code. And so going through this
461. 18:09process has really opened my mind to
462. 18:12we've gotten so used to like the open
463. 18:14chat field. Like if I just type in, the
464. 18:17agent will do good work. And I think we
465. 18:18all have done good work. But then now
466. 18:20I've realized that these agents can help
467. 18:22us solve very, very specific problems
468. 18:25using other agents and by constraining
469. 18:27that work, we can actually get specific
470. 18:29jobs done really efficiently and then
471. 18:32use the general purpose agent to sort of
472. 18:35orchestrate it. So it's really like
473. 18:36changed my mind about how work gets
474. 18:39done.
475. 18:40As you can see here again, it's just a
476. 18:42couple files. It's really not too much.
477. 18:45The adapters to the data sources, um a
478. 18:49couple workflows. In particular, this
479. 18:52bug hunter workflow which just goes
480. 18:54through exactly how we want to hunt
481. 18:56bugs, including how we want to put
482. 18:59together summaries of bug reports and
483. 19:02then some files here in terms of running
484. 19:05the TUI or the CLI. And then as I said,
485. 19:08we have this artifacts folder that gets
486. 19:10updated every time a run happens where I
487. 19:13can click in and actually see exactly
488. 19:17what happened out of a run. So, let's go
489. 19:19and see if this run happened well and
490. 19:24what I can find out. So, now I have the
491. 19:26full context. Here's the investigation
492. 19:28brief and I can go look for it. So, this
493. 19:31is Bug Hunter C7. Let's see if I can
494. 19:35find this one.
495. 19:36Here it is. Here's the investigation
496. 19:38brief on that edit document operations
497. 19:40dropped. I have confirmed evidence. So,
498. 19:43it's saying, "Yes, there was definitely
499. 19:45a Sentry warning. It's impacted 150
500. 19:47users. It's still happening hourly. Um
501. 19:50it's a warning, so it's not an actual
502. 19:52error."
503. 19:54And the Versel logs were unavailable and
504. 19:56so we weren't able to use that data. And
505. 19:58then it found likely root causes. So,
506. 20:01invalid original range or overlapping
507. 20:04original range. And so, it's identified
508. 20:06a couple potential root causes as well
509. 20:09as a blind spot in this particular
510. 20:10function. It's told me exactly where in
511. 20:13the product surface um the issue is and
512. 20:17then how I would actually verify this by
513. 20:20fetching a raw Sentry event to see if
514. 20:23the issues that they've identified are
515. 20:25correct. It's identifying should it um
516. 20:29issue a linear issue and it says, "Yes,
517. 20:31we should definitely make a linear issue
518. 20:32to fix this." And so, this should get
519. 20:34assigned to somebody. And then it
520. 20:36doesn't recommend turning on patch mode
521. 20:39and actually fixing this. So, again,
522. 20:41this is like a very specific outcome I
523. 20:43wanted. I wanted to say like, "What's
524. 20:45all the evidence?
525. 20:46Priority rank the root causes. Make a
526. 20:48suggestion on the next step if we need
527. 20:51to verify this more. Tell me if I need
528. 20:53to assign it to somebody in Linear and
529. 20:55then tell me if you can fix it." And
530. 20:57they're saying, "No, I don't think I can
531. 20:58fix it yet. I need a little bit more
532. 21:00information."
533. 21:02And all of that is built because I have
534. 21:05done this like very specific workflow
535. 21:08and encoded that in
536. 21:11what we're calling a harness, which is
537. 21:12just code around an agent. So, how would
538. 21:16you you build your own harness? I feel
539. 21:18like hopefully you're still with me not
540. 21:20too much of that went over your head.
541. 21:22Just to reiterate, I just identified a
542. 21:25specific workflow. I determined what the
543. 21:29run against the task would look like. I
544. 21:32made very opinionated calls to tools or
545. 21:35data sources, so I didn't just say like
546. 21:37use an MCP, although that could be part
547. 21:39of your harness. But what I did is I
548. 21:41made adapters that made the calls to
549. 21:44these external APIs and tools very
550. 21:45specific. I thought about what the
551. 21:48structured artifacts out of that
552. 21:50workflow might be. I decided what rules
553. 21:53and permissions I wanted to give this
554. 21:55harness and which ones I didn't. I
555. 21:57decided whether I wanted to use Claude
556. 21:59Code or Codex or a model router to
557. 22:02actually run these things. And then I
558. 22:04built a surface to interact with this
559. 22:06agent. So, I built a TUI so I could
560. 22:09actually look and work with this harness
561. 22:12in a way. It could be a TUI, it could be
562. 22:14a CLI, it could be a web app, but I
563. 22:16built some way to interact with this.
564. 22:18So, this is what you need to do.
565. 22:20Identify a workflow. Uh really write it
566. 22:22down on, you know, proverbial paper,
567. 22:24HTML or markdown. Figure out what
568. 22:27sources of data you want and then plug
569. 22:29it all into Claude Code or into Codex as
570. 22:32I did and have it build your own harness
571. 22:35and then test it against real data. So,
572. 22:38that's it. I just I really hope that you
573. 22:41walk away from this realizing that these
574. 22:44mystery terms like harness are not that
575. 22:47mysterious. A harness is simply putting
576. 22:50some structure around how AI works. Yes,
577. 22:53Cursor is like a really complex harness.
578. 22:56Yes, Codex and Claude Code are very
579. 22:58complex coding harnesses. But at the end
580. 23:01of the day, they're code that wraps
581. 23:03these AI agents and these AI calls to
582. 23:05make them more efficient in doing a very
583. 23:08specific job. And so whether you're
584. 23:10doing that in a very prescriptive way
585. 23:12like I just showed where I want to show
586. 23:14you how I triage sentry bugs, do the
587. 23:16investigation and pass it on to the
588. 23:18team, or you're doing it a broad way
589. 23:20like these general purpose coding agents
590. 23:22that just have access to tools and
591. 23:25context and methods that make the coding
592. 23:27workflow better.
593. 23:29That's all harnesses. You can think of
594. 23:31harnesses that you can build. You can
595. 23:33build them in the terminal. You can
596. 23:35build them for CLIs. You can even build
597. 23:37them as web apps. I'm starting to
598. 23:39hypothesize that a wrapper is just a
599. 23:42harness and that is going to upgrade
600. 23:44everything that I've vibe coded over the
601. 23:46last 3 years.
602. 23:48This has been totally a learning
603. 23:50experience for me here on How AI. This
604. 23:52is my very first harness that I've built
605. 23:55live on the show. I hope it's useful for
606. 23:57you. And if you're interested in me
607. 23:59building other and demystifying AI
608. 24:01terms, let me know in the comments.
609. 24:04Thanks for joining How AI.
610. 24:07Thanks so much for watching. If you
611. 24:09enjoyed the show, please like and
612. 24:11subscribe here on YouTube or even
613. 24:13better, leave us a comment with your
614. 24:14thoughts. You can also find this podcast
615. 24:17on Apple Podcasts, Spotify, or your
616. 24:19favorite podcast app. Please consider
617. 24:22leaving us a rating and review which
618. 24:24will help others find the show. You can
619. 24:26see all our episodes and learn more
620. 24:28about the show at howiaiipod.com.
621. 24:32See you next time.