1
07:48:54,398 --> 07:48:56,310
second here.

2
07:48:56,320 --> 07:48:59,030
Any 49ers fans?

3
07:48:59,040 --> 07:49:00,628
>> Okay. You know, that's a better reaction

4
07:49:00,638 --> 07:49:01,910
I've gotten to other events this year

5
07:49:01,920 --> 07:49:03,190
when I've asked, you know, about their

6
07:49:03,200 --> 07:49:04,468
local sports team because I've gotten a

7
07:49:04,478 --> 07:49:05,830
lot of just like muted signs. I have to

8
07:49:05,840 --> 07:49:07,670
hit the, you know, well, wrong audience

9
07:49:07,680 --> 07:49:09,350
for that type of joke. But glad at least

10
07:49:09,360 --> 07:49:15,110
it can hit this time.

11
07:49:15,120 --> 07:49:29,190
Continue

12
07:49:29,200 --> 07:49:46,708
help.

13
07:49:46,718 --> 07:49:51,350
current slide.

14
07:49:51,360 --> 07:49:59,830
That works.

15
07:49:59,840 --> 07:50:01,270
>> Perfect.

16
07:50:01,280 --> 07:50:03,430
>> All right. Good. Well, thanks so much

17
07:50:03,440 --> 07:50:05,110
for waiting everyone. My name is Zach

18
07:50:05,120 --> 07:50:07,270
and I'm an engineer at RAMP. I'm just

19
07:50:07,280 --> 07:50:08,628
really excited to tell you all a bit

20
07:50:08,638 --> 07:50:09,990
about how Ramp has evolved its

21
07:50:10,000 --> 07:50:11,750
engineering culture this year through

22
07:50:11,760 --> 07:50:14,708
the help of our robotic friend Inspect.

23
07:50:14,718 --> 07:50:16,548
As a quick introduction, Inspect is

24
07:50:16,558 --> 07:50:18,708
Ramp's collaborative coding agent. About

25
07:50:18,718 --> 07:50:20,708
a year ago, we decided to take a bet on

26
07:50:20,718 --> 07:50:22,548
building our own coding agent with every

27
07:50:22,558 --> 07:50:24,628
session running and being fully

28
07:50:24,638 --> 07:50:26,788
multiplayer. Inspect can work alongside

29
07:50:26,798 --> 07:50:28,310
you and your team, whether that be an

30
07:50:28,320 --> 07:50:31,510
Inspect itself, Slack, Linear, or any of

31
07:50:31,520 --> 07:50:34,230
our of our integrations. And now it goes

32
07:50:34,240 --> 07:50:36,708
far beyond just writing the code with

33
07:50:36,718 --> 07:50:38,628
the built-in design primitives, review

34
07:50:38,638 --> 07:50:40,868
agents, and verification flows that let

35
07:50:40,878 --> 07:50:43,270
any builder go from prompt to shipping

36
07:50:43,280 --> 07:50:45,510
all inside inspect. Inspect is also

37
07:50:45,520 --> 07:50:46,788
connected with all of our integrations

38
07:50:46,798 --> 07:50:48,708
companywide. So everyone's automatically

39
07:50:48,718 --> 07:50:50,788
connected the to the data and tools that

40
07:50:50,798 --> 07:50:52,788
matter most, whether that be telemetry,

41
07:50:52,798 --> 07:50:54,948
databases, or any of the other tooling

42
07:50:54,958 --> 07:50:57,190
available.

43
07:50:57,200 --> 07:50:59,910
Through this, we now see around 75% of

44
07:50:59,920 --> 07:51:02,948
all of RAMP's code uh being being

45
07:51:02,958 --> 07:51:05,590
written and shipped by inspect.

46
07:51:05,600 --> 07:51:07,350
When we started sharing Inspect publicly

47
07:51:07,360 --> 07:51:08,868
at the start of the year, the technology

48
07:51:08,878 --> 07:51:10,788
itself was the more novel part. Spinning

49
07:51:10,798 --> 07:51:13,270
up sandboxes in under 5 seconds, wiring

50
07:51:13,280 --> 07:51:15,110
up all that internal context, having

51
07:51:15,120 --> 07:51:16,868
agents work together in a multiplayer

52
07:51:16,878 --> 07:51:18,548
setting. People had a lot of questions

53
07:51:18,558 --> 07:51:21,030
about the plumbing. But now with much

54
07:51:21,040 --> 07:51:22,948
broader adoption of these concepts and

55
07:51:22,958 --> 07:51:24,868
Inspect itself now being a product on

56
07:51:24,878 --> 07:51:26,788
the market, the technology isn't nearly

57
07:51:26,798 --> 07:51:29,350
the barrier that it once was. Instead, I

58
07:51:29,360 --> 07:51:31,030
want to talk about the other angle, the

59
07:51:31,040 --> 07:51:32,468
lessons we've learned around building a

60
07:51:32,478 --> 07:51:34,310
culture around Inspect and what you can

61
07:51:34,320 --> 07:51:36,070
take away to inspire the people and

62
07:51:36,080 --> 07:51:39,110
teams that you work with.

63
07:51:39,120 --> 07:51:40,468
Uh I'm not going to talk as much about

64
07:51:40,478 --> 07:51:42,070
those technical aspects of Inspect

65
07:51:42,080 --> 07:51:43,350
today, but if you are curious about

66
07:51:43,360 --> 07:51:45,510
that, you can read our blog posts or

67
07:51:45,520 --> 07:51:47,430
take a look at pragmatic engineers deep

68
07:51:47,440 --> 07:51:48,788
dive that they did with us a few months

69
07:51:48,798 --> 07:51:51,110
ago. You can also follow my boy Dylan

70
07:51:51,120 --> 07:51:53,030
Garcia on Twitter. He's on his social

71
07:51:53,040 --> 07:51:55,430
media grind, but he's also been posting

72
07:51:55,440 --> 07:51:57,750
a ton of really great inspect insights

73
07:51:57,760 --> 07:51:59,430
uh since we've launched it. So, shout

74
07:51:59,440 --> 07:52:02,708
out Dylan. Please give him a follow.

75
07:52:02,718 --> 07:52:04,868
One thing about working at RAMP is that

76
07:52:04,878 --> 07:52:06,548
it really makes you start thinking about

77
07:52:06,558 --> 07:52:08,070
everything through the lens of saving

78
07:52:08,080 --> 07:52:10,628
time and money. And although I think it

79
07:52:10,638 --> 07:52:12,150
is important to understand the systems

80
07:52:12,160 --> 07:52:13,910
that you work with and understand how

81
07:52:13,920 --> 07:52:16,310
your tooling works, if you're a multiund

82
07:52:16,320 --> 07:52:18,948
or multi-,000 person organization and

83
07:52:18,958 --> 07:52:20,948
everyone has to go through say 15

84
07:52:20,958 --> 07:52:23,110
minutes of setup, once you start

85
07:52:23,120 --> 07:52:25,430
multiplying that across an organization,

86
07:52:25,440 --> 07:52:29,910
this is days lost for what is very w.

87
07:52:29,920 --> 07:52:31,510
Then if you consider how fast the

88
07:52:31,520 --> 07:52:33,830
frontier moves here on best practices

89
07:52:33,840 --> 07:52:35,910
and new ways to set up an agent, you

90
07:52:35,920 --> 07:52:37,990
realize this setup cost isn't just one

91
07:52:38,000 --> 07:52:40,788
time. Not to mention these advancements

92
07:52:40,798 --> 07:52:43,270
won't always spread equally. So now you

93
07:52:43,280 --> 07:52:45,030
have an indeterminate amount of time and

94
07:52:45,040 --> 07:52:46,868
effort spent on this maintenance without

95
07:52:46,878 --> 07:52:48,628
an equal spread of knowledge leading to

96
07:52:48,638 --> 07:52:50,230
a disparate outcomes across an

97
07:52:50,240 --> 07:52:52,310
organization. While you can perceive

98
07:52:52,320 --> 07:52:54,070
this as an advantage on individual

99
07:52:54,080 --> 07:52:56,150
level, when you're trying to win, you

100
07:52:56,160 --> 07:52:58,788
need everyone on your agame.

101
07:52:58,798 --> 07:53:01,110
So this was our first lesson. Build

102
07:53:01,120 --> 07:53:03,990
tools that kill information silos. When

103
07:53:04,000 --> 07:53:06,070
we found patterns that worked well or

104
07:53:06,080 --> 07:53:08,230
tools that benefited everyone, we could

105
07:53:08,240 --> 07:53:10,070
push it to our entire organization

106
07:53:10,080 --> 07:53:11,990
immediately. If you want to spend time

107
07:53:12,000 --> 07:53:13,990
tweaking your own setup, that's fine.

108
07:53:14,000 --> 07:53:15,910
But inspect is always available as a

109
07:53:15,920 --> 07:53:18,310
very solid baseline. And what we found

110
07:53:18,320 --> 07:53:20,230
is that we've never needed to mandate

111
07:53:20,240 --> 07:53:22,548
usage of inspect because it existing as

112
07:53:22,558 --> 07:53:24,548
this baseline reduced the need for

113
07:53:24,558 --> 07:53:26,310
people to reach for their own setups all

114
07:53:26,320 --> 07:53:28,948
the time.

115
07:53:28,958 --> 07:53:30,548
Part of what I think makes this powerful

116
07:53:30,558 --> 07:53:32,868
as well is that the reality of this

117
07:53:32,878 --> 07:53:35,270
industry is that it's unlikely that you

118
07:53:35,280 --> 07:53:37,430
as an individual always know the

119
07:53:37,440 --> 07:53:39,190
absolute best way to work with agents

120
07:53:39,200 --> 07:53:41,350
across everything. Even if keeping up

121
07:53:41,360 --> 07:53:43,430
with this is your full-time job, we all

122
07:53:43,440 --> 07:53:46,150
miss things, hold biases, and have blind

123
07:53:46,160 --> 07:53:48,788
spots. So, part of the design of Inspect

124
07:53:48,798 --> 07:53:51,110
is that there's just one single Inspect

125
07:53:51,120 --> 07:53:53,430
for the whole company. When people

126
07:53:53,440 --> 07:53:55,510
figure out better ways to use it or have

127
07:53:55,520 --> 07:53:56,868
ideas for what would make it better for

128
07:53:56,878 --> 07:53:59,110
the work they do, we're able to bring it

129
07:53:59,120 --> 07:54:01,670
to the entire company. The best part of

130
07:54:01,680 --> 07:54:03,590
this is that for most changes, this is

131
07:54:03,600 --> 07:54:05,910
invisible to everyone else. You just

132
07:54:05,920 --> 07:54:07,830
keep using Inspect as you normally would

133
07:54:07,840 --> 07:54:09,430
and benefit from the improvements that

134
07:54:09,440 --> 07:54:12,468
everyone's making over time.

135
07:54:12,478 --> 07:54:13,990
And what we found is that we really

136
07:54:14,000 --> 07:54:16,468
don't have to run these formal sitdown

137
07:54:16,478 --> 07:54:18,388
here's how you use AI effectively

138
07:54:18,398 --> 07:54:20,708
training sessions for Inspect. With

139
07:54:20,718 --> 07:54:22,388
every inspect session being public

140
07:54:22,398 --> 07:54:24,628
within the company by default, it's

141
07:54:24,638 --> 07:54:26,628
actually easier to work in public than

142
07:54:26,638 --> 07:54:29,350
in private. As such, you default to

143
07:54:29,360 --> 07:54:31,190
working in public. And you can learn how

144
07:54:31,200 --> 07:54:33,430
to use it by watching how your teammates

145
07:54:33,440 --> 07:54:34,948
use it in the channels that you're

146
07:54:34,958 --> 07:54:37,110
already in. And your teammates can learn

147
07:54:37,120 --> 07:54:39,590
from you. You can literally ask inspect

148
07:54:39,600 --> 07:54:41,270
to go through their recent sessions and

149
07:54:41,280 --> 07:54:43,030
tell you about how they work, how they

150
07:54:43,040 --> 07:54:44,788
structure prompts, or how they break

151
07:54:44,798 --> 07:54:47,430
down gnarly problems.

152
07:54:47,440 --> 07:54:49,430
I I think this has not just been helpful

153
07:54:49,440 --> 07:54:51,590
for the builders at ramp, but also for

154
07:54:51,600 --> 07:54:54,310
the inspect team itself. We frequently

155
07:54:54,320 --> 07:54:56,788
see the way that people use inspect in

156
07:54:56,798 --> 07:54:59,190
ways that we didn't originally imagine.

157
07:54:59,200 --> 07:55:00,788
This then informs what we should build

158
07:55:00,798 --> 07:55:02,948
towards, whether that should be more

159
07:55:02,958 --> 07:55:05,190
native primitives for common patterns or

160
07:55:05,200 --> 07:55:07,590
where tools and data are missing. All in

161
07:55:07,600 --> 07:55:09,510
all, it creates this self-sustaining

162
07:55:09,520 --> 07:55:11,270
loop where we're able to build towards

163
07:55:11,280 --> 07:55:13,510
what makes Inspect the most successful

164
07:55:13,520 --> 07:55:15,270
by seeing what people are already

165
07:55:15,280 --> 07:55:18,150
achieving with Inspect.

166
07:55:18,160 --> 07:55:19,910
This feeds into the second lesson that

167
07:55:19,920 --> 07:55:21,750
we've learned. It's all about building

168
07:55:21,760 --> 07:55:24,628
trust. I think it's fair to say that

169
07:55:24,638 --> 07:55:26,708
adopting agentic driven coding has been

170
07:55:26,718 --> 07:55:28,628
a massive trust building exercise for

171
07:55:28,638 --> 07:55:30,868
every engineer this year. It's helped

172
07:55:30,878 --> 07:55:32,230
that the models have become more

173
07:55:32,240 --> 07:55:34,070
competent at writing high-quality code.

174
07:55:34,080 --> 07:55:36,230
But there's a large part of this that's

175
07:55:36,240 --> 07:55:37,990
just learning to trust what an agent

176
07:55:38,000 --> 07:55:39,750
builds.

177
07:55:39,760 --> 07:55:41,670
So much of building trust in tools comes

178
07:55:41,680 --> 07:55:43,510
from learning from other people. After

179
07:55:43,520 --> 07:55:44,948
all, that's why most of us are here,

180
07:55:44,958 --> 07:55:47,190
right? You wanted to listen to people

181
07:55:47,200 --> 07:55:49,190
you respect, or at least I hope you

182
07:55:49,200 --> 07:55:51,110
respect, uh, teach you about the things

183
07:55:51,120 --> 07:55:53,430
that they learned. But if we limit our

184
07:55:53,440 --> 07:55:55,350
corpus of knowledge to just those who

185
07:55:55,360 --> 07:55:57,030
find the perverse enjoyment in public

186
07:55:57,040 --> 07:55:59,510
speaking, we're still not getting the

187
07:55:59,520 --> 07:56:02,548
full picture. This democratization of

188
07:56:02,558 --> 07:56:04,388
knowledge sharing is just fascinating to

189
07:56:04,398 --> 07:56:06,628
me as a concept. Instead of grouping

190
07:56:06,638 --> 07:56:08,468
around a select few leaders, can we

191
07:56:08,478 --> 07:56:10,310
build systems that inherently share the

192
07:56:10,320 --> 07:56:13,110
knowledge of everyone? This is what

193
07:56:13,120 --> 07:56:16,628
Inspect has uh built towards.

194
07:56:16,638 --> 07:56:18,230
Every time someone ships something with

195
07:56:18,240 --> 07:56:20,230
Inspect, people can see it right there

196
07:56:20,240 --> 07:56:22,150
on the work you've done. The session

197
07:56:22,160 --> 07:56:24,628
itself is an easy to share. It's a

198
07:56:24,638 --> 07:56:26,468
living artifact of how that change came

199
07:56:26,478 --> 07:56:28,708
to be. When a poll request goes up,

200
07:56:28,718 --> 07:56:30,708
anyone reviewing it doesn't just see the

201
07:56:30,718 --> 07:56:32,948
400line diff dropped over the wall and

202
07:56:32,958 --> 07:56:35,030
hope for the best. They can see the

203
07:56:35,040 --> 07:56:37,990
exact conversation that you had, all the

204
07:56:38,000 --> 07:56:40,388
tools that Inspect used, before and

205
07:56:40,398 --> 07:56:43,030
after screenshots, or even a video

206
07:56:43,040 --> 07:56:44,868
recording of the verification agent

207
07:56:44,878 --> 07:56:46,468
clicking through the app in the browser

208
07:56:46,478 --> 07:56:49,110
to prove its work holds up. It makes it

209
07:56:49,120 --> 07:56:50,948
effortless to show not just what

210
07:56:50,958 --> 07:56:54,548
changed, but why you can trust it.

211
07:56:54,558 --> 07:56:56,310
It also just makes it easier to bring

212
07:56:56,320 --> 07:56:57,830
people you don't work with all the time

213
07:56:57,840 --> 07:57:00,548
into the fold. A fun moment on this one.

214
07:57:00,558 --> 07:57:02,388
It was back in February. One of our

215
07:57:02,398 --> 07:57:04,468
engineers, Andrew, reported a bug in

216
07:57:04,478 --> 07:57:07,110
Inspect. So, naturally, I reached for

217
07:57:07,120 --> 07:57:09,910
Inspect and asked to investigate. What's

218
07:57:09,920 --> 07:57:11,990
funny here is that although Inspect had

219
07:57:12,000 --> 07:57:14,070
access to data dog for a while, we'd

220
07:57:14,080 --> 07:57:15,670
only just finished instrumenting inspect

221
07:57:15,680 --> 07:57:18,070
itself. I mean, inspect was primarily

222
07:57:18,080 --> 07:57:20,310
built by front-end e- product engineers.

223
07:57:20,320 --> 07:57:22,070
So, tracing is a little bit of a foreign

224
07:57:22,080 --> 07:57:24,468
concept for us. But in any case, it was

225
07:57:24,478 --> 07:57:26,150
a great time for us to put it to the

226
07:57:26,160 --> 07:57:30,150
test for inspect. And inspect says

227
07:57:30,160 --> 07:57:31,910
the instrumentation was causing the

228
07:57:31,920 --> 07:57:33,830
error. Yes, the very instrumentation

229
07:57:33,840 --> 07:57:35,670
that was letting us debug this error was

230
07:57:35,680 --> 07:57:38,708
actually causing this error. So, Dylan

231
07:57:38,718 --> 07:57:40,788
and I were not exactly convinced by the

232
07:57:40,798 --> 07:57:43,350
robot here. And Andrew is just like,

233
07:57:43,360 --> 07:57:45,590
whatever, I'll use cloud code. And and

234
07:57:45,600 --> 07:57:47,590
while we don't have a mandate to use

235
07:57:47,600 --> 07:57:50,230
inspect, I mean, personally, these were

236
07:57:50,240 --> 07:57:52,868
not the words I wanted to hear.

237
07:57:52,878 --> 07:57:55,270
So, nobody here is really trusting

238
07:57:55,280 --> 07:57:57,990
anybody. Dylan and I don't trust the

239
07:57:58,000 --> 07:58:00,628
robots explanation. Andrew doesn't trust

240
07:58:00,638 --> 07:58:02,548
us to build something that works. And

241
07:58:02,558 --> 07:58:05,270
inspect, well, I I I don't know if

242
07:58:05,280 --> 07:58:06,948
Inspect trusts anyone, but that's okay.

243
07:58:06,958 --> 07:58:09,590
That's okay. But because this is

244
07:58:09,600 --> 07:58:12,468
happening together in public, we keep

245
07:58:12,478 --> 07:58:14,708
proddding at it. Dylan asks inspect to

246
07:58:14,718 --> 07:58:16,070
look through the code of the telemetry

247
07:58:16,080 --> 07:58:18,230
library using and see if this was just

248
07:58:18,240 --> 07:58:19,510
something that was happening before

249
07:58:19,520 --> 07:58:21,990
without visibility because surely by now

250
07:58:22,000 --> 07:58:23,510
coding has been solved and there can't

251
07:58:23,520 --> 07:58:26,230
be bugs that we introduced and this time

252
07:58:26,240 --> 07:58:27,670
it comes back with something a bit more

253
07:58:27,680 --> 07:58:30,388
specific. The hotel library calls fetch

254
07:58:30,398 --> 07:58:31,990
but it never reads or cancels the

255
07:58:32,000 --> 07:58:33,750
response body. This was just a

256
07:58:33,760 --> 07:58:35,510
cloudflare platform limitation at the

257
07:58:35,520 --> 07:58:37,350
time albeit happening in a library

258
07:58:37,360 --> 07:58:39,510
specifically made for cloudflare workers

259
07:58:39,520 --> 07:58:41,270
but that happens. This library was

260
07:58:41,280 --> 07:58:44,310
written pre-agi after all. At this

261
07:58:44,320 --> 07:58:46,388
point, we ask inspect to make a patch.

262
07:58:46,398 --> 07:58:48,468
It updates the exporter to explicitly

263
07:58:48,478 --> 07:58:51,670
cancel the response body. We grumbled in

264
07:58:51,680 --> 07:58:53,190
the thread a bit while it's deploying.

265
07:58:53,200 --> 07:58:54,948
But we see after it deployed, there

266
07:58:54,958 --> 07:58:57,270
haven't been any errors since. So great,

267
07:58:57,280 --> 07:58:59,350
inspect fixed the bug, which at this

268
07:58:59,360 --> 07:59:01,670
point it's nothing special in itself.

269
07:59:01,680 --> 07:59:03,270
But there was a much more important

270
07:59:03,280 --> 07:59:06,708
story playing out in this thread.

271
07:59:06,718 --> 07:59:08,628
Before inspect, a report like this might

272
07:59:08,638 --> 07:59:10,868
have been acknowledged. a ticket filed

273
07:59:10,878 --> 07:59:12,310
and then one of us would have gone away

274
07:59:12,320 --> 07:59:14,230
to debug it. Andrew might have received

275
07:59:14,240 --> 07:59:16,468
a message later saying it was fixed, but

276
07:59:16,478 --> 07:59:17,670
the interesting work would have been

277
07:59:17,680 --> 07:59:20,708
done in private. Now though, it's easier

278
07:59:20,718 --> 07:59:23,910
to do all of this in public.

279
07:59:23,920 --> 07:59:25,910
And as such, Andrew naturally starts

280
07:59:25,920 --> 07:59:27,510
becoming engaged, wanting to chime in

281
07:59:27,520 --> 07:59:29,510
with his thoughts and experiences. And

282
07:59:29,520 --> 07:59:31,910
then in that same thread, this leads to

283
07:59:31,920 --> 07:59:33,510
ideas coming up about how Inspect could

284
07:59:33,520 --> 07:59:35,750
power other work. And none of this

285
07:59:35,760 --> 07:59:37,590
required any proddding or pressure to

286
07:59:37,600 --> 07:59:39,510
work together. instead it was just like

287
07:59:39,520 --> 07:59:41,430
oh these guys are doing something cool

288
07:59:41,440 --> 07:59:45,030
and we could work on it together.

289
07:59:45,040 --> 07:59:47,590
So in the span of an hour we go from a

290
07:59:47,600 --> 07:59:49,590
skeptical user with a broken user

291
07:59:49,600 --> 07:59:51,350
experience to an engaged colleague

292
07:59:51,360 --> 07:59:53,510
wanting to build together and a more

293
07:59:53,520 --> 07:59:55,670
stable inspect.

294
07:59:55,680 --> 07:59:57,750
So once you have a tool that gives

295
07:59:57,760 --> 07:59:59,830
everyone the best setup on day one with

296
07:59:59,840 --> 08:00:02,230
no information silos and you've built a

297
08:00:02,240 --> 08:00:04,070
system of trust that brings innovation

298
08:00:04,080 --> 08:00:07,350
together, what happens next? This is the

299
08:00:07,360 --> 08:00:09,510
last lesson I want to share. I think the

300
08:00:09,520 --> 08:00:11,350
only thing left is to think bigger

301
08:00:11,360 --> 08:00:13,430
together.

302
08:00:13,440 --> 08:00:15,030
Something that's pained me this year uh

303
08:00:15,040 --> 08:00:16,788
is how much reverence has been built for

304
08:00:16,798 --> 08:00:19,590
the lone wolf operator. It's great that

305
08:00:19,600 --> 08:00:21,430
we can all do more individually with the

306
08:00:21,440 --> 08:00:23,750
help of AI, but this can't be the

307
08:00:23,760 --> 08:00:26,548
endgame in my opinion. If we don't build

308
08:00:26,558 --> 08:00:29,110
outwards, collaborate with other people,

309
08:00:29,120 --> 08:00:31,030
and continue to share the progress and

310
08:00:31,040 --> 08:00:33,270
learnings we make, I think this is all

311
08:00:33,280 --> 08:00:35,350
going to collapse in on itself. We'll

312
08:00:35,360 --> 08:00:37,030
only consider ourselves as our target

313
08:00:37,040 --> 08:00:39,110
audience and we won't grow through the

314
08:00:39,120 --> 08:00:41,190
consideration of others perspectives and

315
08:00:41,200 --> 08:00:42,868
experiences.

316
08:00:42,878 --> 08:00:44,150
This is not to say I'm all doom and

317
08:00:44,160 --> 08:00:45,830
gloom about this. The fact we're all

318
08:00:45,840 --> 08:00:47,350
here hopefully is a testament to the

319
08:00:47,360 --> 08:00:49,750
value of connection. More importantly

320
08:00:49,760 --> 08:00:51,510
though, I think we can shape our tools

321
08:00:51,520 --> 08:00:53,350
to enable these connections to flourish

322
08:00:53,360 --> 08:00:56,310
even stronger than before.

323
08:00:56,320 --> 08:00:57,990
That story I told about Andrew has

324
08:00:58,000 --> 08:00:59,990
played out so many times within ramp.

325
08:01:00,000 --> 08:01:01,430
It's probably the most inspiring thing

326
08:01:01,440 --> 08:01:04,150
about Inspect to me, honestly. Sure,

327
08:01:04,160 --> 08:01:06,388
it's a coding agent, but remarkably,

328
08:01:06,398 --> 08:01:08,070
it's become part of the culture.

329
08:01:08,080 --> 08:01:10,548
Everyone at RAMP knows its face. We

330
08:01:10,558 --> 08:01:13,110
dress it up on holidays. We throw in

331
08:01:13,120 --> 08:01:15,910
Easter eggs over time. And we evolve it

332
08:01:15,920 --> 08:01:18,788
as RMP itself has evolved. Not all

333
08:01:18,798 --> 08:01:20,310
iterations have been perfect, as any

334
08:01:20,320 --> 08:01:22,468
RAMP engineer could tell you. But I do

335
08:01:22,478 --> 08:01:26,388
think every iteration has been together.

336
08:01:26,398 --> 08:01:28,468
Going back to earlier this year, Inspect

337
08:01:28,478 --> 08:01:29,590
didn't have nearly the amount of

338
08:01:29,600 --> 08:01:31,670
functionality it has today. A great

339
08:01:31,680 --> 08:01:34,708
example is its code review agent. Before

340
08:01:34,718 --> 08:01:36,310
this, we were using various

341
08:01:36,320 --> 08:01:37,990
off-the-shelf code review products,

342
08:01:38,000 --> 08:01:39,750
frequently rotating between them to find

343
08:01:39,760 --> 08:01:42,150
a good fit. At some point though, a

344
08:01:42,160 --> 08:01:43,910
couple engineers, Daniel and Franel,

345
08:01:43,920 --> 08:01:45,750
realized that we could probably build it

346
08:01:45,760 --> 08:01:47,990
ourselves as Inspect already had

347
08:01:48,000 --> 08:01:49,750
automation triggers that fired from a

348
08:01:49,760 --> 08:01:52,070
variety of sources, including when a

349
08:01:52,080 --> 08:01:54,708
pull request was opened. With just what

350
08:01:54,718 --> 08:01:56,468
we built already, they got to work

351
08:01:56,478 --> 08:01:58,548
building another beloved agent at RAMP,

352
08:01:58,558 --> 08:02:00,868
review buddy. And what was so cool about

353
08:02:00,878 --> 08:02:02,948
this is that much like Inspec's own

354
08:02:02,958 --> 08:02:05,430
origins as a side project, these

355
08:02:05,440 --> 08:02:07,350
engineers were working on this. They

356
08:02:07,360 --> 08:02:08,788
didn't need to ask for permission to

357
08:02:08,798 --> 08:02:10,468
build it. They just challenged

358
08:02:10,478 --> 08:02:12,310
themselves to build something better,

359
08:02:12,320 --> 08:02:14,548
used inspect to move faster, and quickly

360
08:02:14,558 --> 08:02:16,468
built an agent that was better for all

361
08:02:16,478 --> 08:02:18,548
of ramp.

362
08:02:18,558 --> 08:02:20,388
This really did force us working on

363
08:02:20,398 --> 08:02:22,150
Inspect to think about Inspect

364
08:02:22,160 --> 08:02:24,868
differently. At that point in time, we

365
08:02:24,878 --> 08:02:26,548
were pretty okay with Inspect quote

366
08:02:26,558 --> 08:02:29,270
unquote just being a coding agent. But

367
08:02:29,280 --> 08:02:31,350
once we saw what others were pushing it

368
08:02:31,360 --> 08:02:34,070
to do, it got us thinking about how

369
08:02:34,080 --> 08:02:36,788
Inspect could become full circle. If

370
08:02:36,798 --> 08:02:38,788
this push didn't happen naturally, I

371
08:02:38,798 --> 08:02:40,230
think it would have taken the Inspect

372
08:02:40,240 --> 08:02:44,150
team a lot longer to get there.

373
08:02:44,160 --> 08:02:45,830
This is what I mean about thinking

374
08:02:45,840 --> 08:02:48,070
bigger together. Especially with the

375
08:02:48,080 --> 08:02:49,430
continued improvements in model

376
08:02:49,440 --> 08:02:51,590
performance, now is the time to push

377
08:02:51,600 --> 08:02:53,830
yourself and the people around you to

378
08:02:53,840 --> 08:02:56,230
think bigger. I encourage all of you to

379
08:02:56,240 --> 08:02:57,990
take some time in this coming week to

380
08:02:58,000 --> 08:03:00,150
ask everyone that you work with, what

381
08:03:00,160 --> 08:03:02,388
have we been waiting to do that we

382
08:03:02,398 --> 08:03:04,868
haven't done yet. And once you found

383
08:03:04,878 --> 08:03:06,708
those things, try to build them

384
08:03:06,718 --> 08:03:08,948
together. There's really not an excuse

385
08:03:08,958 --> 08:03:11,270
to not just try anymore. You have access

386
08:03:11,280 --> 08:03:13,830
to the same models as everyone else. But

387
08:03:13,840 --> 08:03:15,910
you have to start somewhere. And I

388
08:03:15,920 --> 08:03:17,990
believe that there is so much potential

389
08:03:18,000 --> 08:03:19,910
in unlocking wins, not just for

390
08:03:19,920 --> 08:03:22,628
yourself, but for the people around you.

391
08:03:22,638 --> 08:03:24,548
Inspect itself was born this way. not

392
08:03:24,558 --> 08:03:26,230
only as something that I wanted to build

393
08:03:26,240 --> 08:03:28,388
and Jason wanted to build, but out of a

394
08:03:28,398 --> 08:03:30,628
want to build something for a team that

395
08:03:30,638 --> 08:03:33,190
I truly love and truly respect that

396
08:03:33,200 --> 08:03:35,030
would just make their jobs a little more

397
08:03:35,040 --> 08:03:36,948
enjoyable.

398
08:03:36,958 --> 08:03:38,628
Now, if you'd like to use a coding agent

399
08:03:38,638 --> 08:03:40,150
that has embodied all these lessons that

400
08:03:40,160 --> 08:03:41,510
we've learned, I'm happy to say that

401
08:03:41,520 --> 08:03:43,350
Inspect is now available outside of

402
08:03:43,360 --> 08:03:47,030
RAMP. If you go to inspect.com/ieie,

403
08:03:47,040 --> 08:03:48,708
we'll prioritize you and your team on

404
08:03:48,718 --> 08:03:49,990
the wait list. We're actually moving

405
08:03:50,000 --> 08:03:51,350
through this weight list pretty quickly,

406
08:03:51,360 --> 08:03:53,750
so expect to hear from us soon. But

407
08:03:53,760 --> 08:03:55,510
otherwise, I hope you've enjoyed hearing

408
08:03:55,520 --> 08:03:57,430
about our robotic friend, Inspect. If

409
08:03:57,440 --> 08:03:58,948
these topics interest you, feel free to

410
08:03:58,958 --> 08:04:01,110
reach out here or an X. But otherwise,

411
08:04:01,120 --> 08:04:03,680
thanks so much for listening.