1
00:00:00,080 --> 00:00:02,440
OK, so left. 
Let's unpack this right out of 

2
00:00:02,440 --> 00:00:05,720
the gate. 
For like the last two months or 

3
00:00:05,720 --> 00:00:09,880
so, developers worldwide have 
been completely convinced that 

4
00:00:09,880 --> 00:00:12,800
their AI coding assistants were 
quietly losing their minds. 

5
00:00:12,800 --> 00:00:14,880
Oh, totally. 
It was a massive panic, right? 

6
00:00:15,000 --> 00:00:17,840
Like if you've been relying on 
Claude or Copilot to push code 

7
00:00:17,840 --> 00:00:19,920
lately, you probably felt it 
yourself. 

8
00:00:19,920 --> 00:00:22,120
The code just, it wasn't landing
right. 

9
00:00:22,120 --> 00:00:23,960
It felt sloppy. 
Yeah, it felt broken. 

10
00:00:24,040 --> 00:00:26,680
Exactly broken. 
But it turns out the AI was 

11
00:00:26,680 --> 00:00:30,160
actually completely fine, which 
is while it was the invisible 

12
00:00:30,160 --> 00:00:34,320
scaffolding holding the entire 
system together that was quietly

13
00:00:34,320 --> 00:00:37,600
falling apart. 
So today we are pulling from a 

14
00:00:37,600 --> 00:00:42,080
really fascinating stack of late
April 2026 sources. 

15
00:00:42,080 --> 00:00:45,560
We've got developer blogs, deep 
dive form threads, breaking tech

16
00:00:45,560 --> 00:00:48,160
news, and some podcast 
transcripts from the actual 

17
00:00:48,160 --> 00:00:49,840
front lines of enterprise 
engineering. 

18
00:00:49,920 --> 00:00:51,160
The real boots on the ground 
stuff. 

19
00:00:51,480 --> 00:00:53,360
Right. 
And our mission for this deep 

20
00:00:53,360 --> 00:00:56,400
dive is to just cut through all 
that noise and chart the 

21
00:00:56,400 --> 00:00:58,000
dramatic shift that's happening 
right now. 

22
00:00:58,000 --> 00:01:01,360
Like literally early as we 
speak, we're watching AI move 

23
00:01:01,360 --> 00:01:04,560
from being this, you know, 
helpful little typing assistant 

24
00:01:04,560 --> 00:01:08,840
living in your local code editor
to becoming a fully autonomous 

25
00:01:08,840 --> 00:01:11,760
cloud native agent operating in 
a massive enterprise scale. 

26
00:01:12,040 --> 00:01:14,600
And what's really fascinating 
here is that we are watching the

27
00:01:14,600 --> 00:01:18,080
architecture of software 
development change in real time.

28
00:01:18,080 --> 00:01:21,840
I mean the transition from 
having an AI just write a simple

29
00:01:21,840 --> 00:01:26,080
Python script to having an AI 
independently execute, test, and

30
00:01:26,080 --> 00:01:28,760
secure entire cloud 
infrastructures. 

31
00:01:29,360 --> 00:01:31,960
It's violent, Violent. 
Is a good word for it, honestly.

32
00:01:31,960 --> 00:01:33,920
It is. 
It's breaking the tools we rely 

33
00:01:33,920 --> 00:01:35,440
on. 
We're seeing these incredible 

34
00:01:35,440 --> 00:01:38,680
leaps and capability, but as our
sources show, the industry is 

35
00:01:38,680 --> 00:01:41,280
really struggling to build the 
plane while it's already in 

36
00:01:41,280 --> 00:01:43,360
orbit. 
Which I mean, that brings us 

37
00:01:43,360 --> 00:01:45,800
right back to this massive 
phantom bug that had the whole 

38
00:01:45,800 --> 00:01:47,440
developer community pulling its.
Hair. 

39
00:01:47,440 --> 00:01:49,640
Oh yeah, the Reddit threads for 
Wild. 

40
00:01:49,920 --> 00:01:52,200
They were blowing up. 
People on GitHub were literally 

41
00:01:52,200 --> 00:01:55,480
asking if Claude code had 
suffered some sort of silent 

42
00:01:55,480 --> 00:01:58,800
lobotomy. 
But then Simon Willison, who by 

43
00:01:58,800 --> 00:02:01,960
the way has been rigorously 
tracking this stuff, he 

44
00:02:01,960 --> 00:02:04,520
confirmed that the degradation 
was real, but the community was 

45
00:02:04,520 --> 00:02:06,080
basically diagnosing the wrong 
disease. 

46
00:02:06,280 --> 00:02:08,080
Right. 
And Willison's investigation is 

47
00:02:08,080 --> 00:02:11,039
just, it's a master class in 
understanding how these AI 

48
00:02:11,039 --> 00:02:12,840
products are actually built 
under the hood. 

49
00:02:13,360 --> 00:02:16,920
He discovered 3 distinct 
critical bugs that have been 

50
00:02:16,920 --> 00:02:21,480
introduced over a 2 month 
period, but the kicker is none 

51
00:02:21,480 --> 00:02:24,240
of those bugs were in the 
underlying clawed model itself. 

52
00:02:24,240 --> 00:02:25,560
Wait, none of them. 
None of them. 

53
00:02:25,560 --> 00:02:28,400
They were all located in the 
execution harness because you 

54
00:02:28,400 --> 00:02:31,360
know when you use a tool like 
Clawed Code, you aren't just 

55
00:02:31,360 --> 00:02:34,080
talking to a raw neural network.
Right, there's a whole system. 

56
00:02:34,200 --> 00:02:36,760
Exactly. 
There's this massive layer of 

57
00:02:36,760 --> 00:02:40,720
traditional handwritten code 
acting as a repper, and this 

58
00:02:40,720 --> 00:02:45,000
wrapper takes the AI's plain 
text output, parses it, often 

59
00:02:45,000 --> 00:02:48,080
using incredibly fragile regular
expressions, by the way, and 

60
00:02:48,080 --> 00:02:51,400
then translates it into actual 
command line execution or file 

61
00:02:51,400 --> 00:02:54,280
modifications. 
OK, so it's it's exactly like 

62
00:02:54,280 --> 00:02:58,960
hiring a Michelin star chef, but
the waiter keeps dropping the 

63
00:02:58,960 --> 00:03:01,800
plates on the way to the table. 
That is the perfect analogy, 

64
00:03:01,800 --> 00:03:04,080
yes. 
Right, like you blame the chef 

65
00:03:04,080 --> 00:03:07,480
because your dinner is a mess on
the floor, but the food was 

66
00:03:07,480 --> 00:03:10,160
cooked perfectly. 
The delivery mechanism is what's

67
00:03:10,160 --> 00:03:11,080
completely. 
Broken. 

68
00:03:11,080 --> 00:03:14,520
Exactly. 
The parser might be looking for,

69
00:03:14,520 --> 00:03:19,320
say, a specific markdown block, 
and if the AI adds just one 

70
00:03:19,320 --> 00:03:23,920
weird space, the rejects fails, 
the code doesn't execute, and 

71
00:03:23,920 --> 00:03:26,480
the human developer just 
assumes, oh, the AI has 

72
00:03:26,480 --> 00:03:27,760
degraded. 
Wow. 

73
00:03:27,960 --> 00:03:30,120
And that distinction, I mean, 
that changes everything about 

74
00:03:30,120 --> 00:03:32,760
how practitioners need to 
approach these tools, right? 

75
00:03:32,760 --> 00:03:34,760
It fundamentally shifts the 
responsibility, yeah. 

76
00:03:35,040 --> 00:03:37,240
But the thing that makes this 
situation particularly 

77
00:03:37,240 --> 00:03:39,920
frustrating, at least for the 
community, was anthropic 

78
00:03:39,920 --> 00:03:42,320
silence. 
Like there was no official 

79
00:03:42,320 --> 00:03:44,240
change log admitting the harness
was broken. 

80
00:03:44,240 --> 00:03:46,200
Nothing. 
Zero communication. 

81
00:03:46,200 --> 00:03:47,280
The. 
Community basically had to 

82
00:03:47,280 --> 00:03:50,080
Crowdsource the detection of 
these bugs across different 

83
00:03:50,080 --> 00:03:52,600
forums just to prove they 
weren't imagining things and. 

84
00:03:52,600 --> 00:03:54,400
That's a real problem. 
But hold on. 

85
00:03:54,600 --> 00:03:57,680
If Anthropic is silently 
deploying broken wrappers and 

86
00:03:57,680 --> 00:04:00,640
just, you know, not telling 
anyone, how do developers 

87
00:04:00,640 --> 00:04:04,240
actually trust these platforms 
with mission critical work? 

88
00:04:04,440 --> 00:04:06,320
Well, that's the $1,000,000 
question. 

89
00:04:06,320 --> 00:04:10,120
Or the $100 question, especially
when we saw that bizarre pricing

90
00:04:10,120 --> 00:04:13,880
leak on April 22nd? 
Oh right, the Internet Archive 

91
00:04:13,880 --> 00:04:16,480
catch. 
Yeah, where Anthropic 

92
00:04:16,480 --> 00:04:20,920
accidentally published $100 a 
month enterprise add on for 

93
00:04:20,920 --> 00:04:23,280
Claude code. 
They pulled it down within 

94
00:04:23,280 --> 00:04:24,920
hours, but the archive caught 
it. 

95
00:04:25,360 --> 00:04:28,920
And then this is the crazy part.
GitHub quietly adjusted their 

96
00:04:28,920 --> 00:04:31,480
Copilot individual pricing the 
exact same day. 

97
00:04:31,600 --> 00:04:34,800
The timing is not a coincidence.
Obviously not, but if they are 

98
00:04:34,800 --> 00:04:37,320
gearing up to charge $100 a 
month for these agents, 

99
00:04:37,720 --> 00:04:40,160
shouldn't they be providing, I 
don't know, rock solid 

100
00:04:40,160 --> 00:04:42,160
reliability? 
I mean, they absolutely should, 

101
00:04:42,160 --> 00:04:44,640
but the current reality of the 
market is that the technology is

102
00:04:44,640 --> 00:04:48,080
just moving too fast for 
traditional quality assurance. 

103
00:04:48,400 --> 00:04:50,920
That $100 price tag is highly 
revealing though. 

104
00:04:51,240 --> 00:04:54,440
It signals that both Anthropic 
and GitHub are moving away from 

105
00:04:54,440 --> 00:04:56,480
selling a simple auto complete 
feature. 

106
00:04:56,800 --> 00:04:59,640
They are preparing to sell 
autonomous software engineering 

107
00:04:59,640 --> 00:05:01,920
labor. 
Right, which is a huge shift in 

108
00:05:01,920 --> 00:05:03,080
value. 
Massive. 

109
00:05:03,160 --> 00:05:07,120
But to your point about trust, 
you cannot outsource your trust 

110
00:05:07,120 --> 00:05:10,320
to the vendor anymore. 
This is exactly why any serious 

111
00:05:10,320 --> 00:05:13,960
developer right now must 
maintain a personal evil set. 

112
00:05:13,960 --> 00:05:17,000
Okay, stop right there, because 
if I am paying an enterprise 

113
00:05:17,000 --> 00:05:20,400
premium, why am I the one 
responsible for queuing the AI? 

114
00:05:20,400 --> 00:05:22,800
It sounds backward, I know. 
It really does. 

115
00:05:22,920 --> 00:05:25,680
What does an evil set even look 
like for, you know, person? 

116
00:05:25,680 --> 00:05:28,080
Listening to this right now. 
Well, the reason you have to do 

117
00:05:28,080 --> 00:05:31,720
it is because Anthropic doesn't 
know your specific proprietary 

118
00:05:31,720 --> 00:05:33,000
code base. 
OK, for a point. 

119
00:05:33,240 --> 00:05:35,280
Right? 
So an evil set is basically a 

120
00:05:35,280 --> 00:05:39,480
standardized automated batch of 
coding tasks specific to your 

121
00:05:39,480 --> 00:05:41,480
environment that you run every 
single week. 

122
00:05:41,600 --> 00:05:43,400
Got it. 
So you might have a test that 

123
00:05:43,400 --> 00:05:47,400
asks the AI to refactor a 
specific API endpoint, or to 

124
00:05:47,400 --> 00:05:50,400
write a database migration using
your company's weird custom 

125
00:05:50,400 --> 00:05:53,080
naming conventions, and then you
score its performance. 

126
00:05:53,360 --> 00:05:55,720
Oh, I see. 
So if you run your evil set on a

127
00:05:55,720 --> 00:06:00,520
Monday and it scores A-95, but 
then on Friday it scores a 60, 

128
00:06:00,880 --> 00:06:03,080
you have caught the quality 
drift yourself. 

129
00:06:03,120 --> 00:06:05,520
You isolate whether the chef or 
the waiter is broken. 

130
00:06:05,520 --> 00:06:08,760
Exactly, and you adjust your 
workflow before it breaks your 

131
00:06:08,760 --> 00:06:10,640
production servers. 
So you essentially have to 

132
00:06:10,640 --> 00:06:14,240
become a manager who constantly 
test their employees. 

133
00:06:14,320 --> 00:06:17,200
That's the reality of it now. 
That actually makes a lot of 

134
00:06:17,200 --> 00:06:20,760
sense, especially if the scope 
of what these tools are doing is

135
00:06:20,760 --> 00:06:24,760
expanding so fast. 
Because the next massive shift 

136
00:06:25,120 --> 00:06:28,600
our sources highlight shows that
we aren't just paying for text 

137
00:06:28,600 --> 00:06:30,040
generation anymore. 
Not at all. 

138
00:06:30,040 --> 00:06:32,600
We are paying for the 
environment the AI lives in. 

139
00:06:32,800 --> 00:06:35,480
Like Boris Cherney who is the 
actual creator of Claude Code. 

140
00:06:35,480 --> 00:06:39,760
He posted some guidance on X on 
April 24th that basically 

141
00:06:39,960 --> 00:06:42,240
completely up ends how 
developers have worked for 

142
00:06:42,240 --> 00:06:44,640
decades. 
Yeah, Cherney's thesis is really

143
00:06:44,640 --> 00:06:47,200
addressing A fundamental 
limitation in how we currently 

144
00:06:47,200 --> 00:06:49,840
deploy AI. 
He argued that cloud native code

145
00:06:49,840 --> 00:06:53,120
is essentially meaningless 
without the environment it is to

146
00:06:53,120 --> 00:06:54,800
run in. 
Meaningless. 

147
00:06:54,960 --> 00:06:57,360
That's a strong word, it is, but
it's true. 

148
00:06:57,960 --> 00:06:59,880
When we talk about cloud native 
code, we're talking about 

149
00:06:59,880 --> 00:07:03,320
Terraform scripts that provision
servers, Kubernetes 

150
00:07:03,320 --> 00:07:07,640
configurations that orchestrate 
containers, or IAM policies that

151
00:07:07,640 --> 00:07:09,280
dictate security permissions. 
OK. 

152
00:07:09,360 --> 00:07:13,400
Let me jump in here because it 
an IAM policy identity and 

153
00:07:13,400 --> 00:07:15,600
access management is like the 
perfect example for this. 

154
00:07:15,600 --> 00:07:20,240
No, for it an AI can write an 
IAM Jason file that is 

155
00:07:20,240 --> 00:07:22,120
syntactically flawless. 
Like it? 

156
00:07:22,360 --> 00:07:24,360
It looks beautiful in your local
text editor. 

157
00:07:24,360 --> 00:07:25,760
Perfect formatting. 
Right, right. 

158
00:07:25,760 --> 00:07:29,280
But when you push it to AWS, it 
completely breaks because 

159
00:07:29,280 --> 00:07:33,160
there's some hidden circular 
dependency in your specific 

160
00:07:33,160 --> 00:07:34,760
cloud architecture. 
Exactly. 

161
00:07:34,880 --> 00:07:36,960
It's like testing cloud 
infrastructure in a local mock 

162
00:07:36,960 --> 00:07:39,960
environment on your laptop is 
like trying to test a submarine 

163
00:07:39,960 --> 00:07:41,640
in your bathtub. 
That is a great visual. 

164
00:07:41,960 --> 00:07:44,360
You need the real ocean. 
Your bathtub doesn't have water 

165
00:07:44,360 --> 00:07:47,520
pressure, and your laptop 
doesn't have the complex network

166
00:07:47,520 --> 00:07:50,560
latencies or the bizarre 
permission errors of a live 

167
00:07:50,560 --> 00:07:53,840
production cloud environment. 
And that submarine analogy 

168
00:07:53,840 --> 00:07:56,640
illustrates the exact friction 
point the industry is hitting 

169
00:07:56,640 --> 00:07:58,920
right now. 
The core product of an 

170
00:07:58,920 --> 00:08:01,960
autonomous coding agent isn't 
the initial code generation 

171
00:08:01,960 --> 00:08:04,680
anymore. 
The real value is the agentic 

172
00:08:04,680 --> 00:08:07,800
feedback loop. 
And that loop requires 4 steps, 

173
00:08:07,800 --> 00:08:10,760
right Execute, observe and 
correct. 

174
00:08:10,960 --> 00:08:12,360
OK, right? 
Execute. 

175
00:08:12,360 --> 00:08:13,440
Observe. 
Correct. 

176
00:08:13,520 --> 00:08:16,280
Yep. 
So when an AI writes A terraform

177
00:08:16,280 --> 00:08:18,400
script locally, it stops at 
right? 

178
00:08:18,720 --> 00:08:22,000
It literally cannot observe the 
cloud provider kicking back a 

179
00:08:22,000 --> 00:08:25,000
four O 3 forbidden error. 
And if it cannot observe the 

180
00:08:25,000 --> 00:08:27,160
error, it cannot correct its own
work. 

181
00:08:27,160 --> 00:08:29,520
Oh wow, and developers are 
taking this seriously? 

182
00:08:29,520 --> 00:08:32,440
I mean look at root code. 
Yeah, the open source extension.

183
00:08:32,440 --> 00:08:35,120
Right, this was an incredibly 
popular extension that lived 

184
00:08:35,120 --> 00:08:38,520
inside VS Code and they just 
announced they are shutting down

185
00:08:38,520 --> 00:08:42,360
their IDE extension entirely to 
pivot to a cloud based agent. 

186
00:08:42,840 --> 00:08:45,960
They explicitly declared and I 
quote ID. 

187
00:08:45,960 --> 00:08:48,440
ES aren't the future of coding. 
Which is a massive stomach. 

188
00:08:48,440 --> 00:08:51,640
It's huge abandoning the IDE. 
That is where software engineers

189
00:08:51,640 --> 00:08:53,360
have basically lived for 30 
years. 

190
00:08:53,720 --> 00:08:56,440
Is the industry really ready to 
just walk away from the local 

191
00:08:56,440 --> 00:08:58,680
development environment? 
I mean, the transition won't 

192
00:08:58,680 --> 00:09:02,040
happen overnight obviously, but 
structurally it has to happen. 

193
00:09:02,040 --> 00:09:04,440
Why though? 
Just for the feedback loop. 

194
00:09:04,520 --> 00:09:08,320
Well, if we connect this to the 
bigger picture, giving an AI 

195
00:09:08,360 --> 00:09:11,880
just file right permissions on 
your MacBook heavily restricts 

196
00:09:11,880 --> 00:09:13,760
its utility. 
You're cutting its potential in 

197
00:09:13,760 --> 00:09:14,680
half. 
OK, Yeah. 

198
00:09:14,720 --> 00:09:19,160
To unleash its full capability, 
it needs execution permissions 

199
00:09:19,160 --> 00:09:23,000
and a dedicated sandbox dev 
cluster in the cloud. 

200
00:09:23,560 --> 00:09:27,120
You provision a rectica of your 
cloud environment, give the AI 

201
00:09:27,120 --> 00:09:30,280
the keys to deploy code there, 
and you just let it run into 

202
00:09:30,280 --> 00:09:31,360
walls. 
Let it crash. 

203
00:09:31,440 --> 00:09:34,080
Exactly. 
It reads the crash logs, it 

204
00:09:34,080 --> 00:09:35,880
rewrites the code and tries 
again. 

205
00:09:36,120 --> 00:09:39,880
We are moving from tools that 
live inside your editor to these

206
00:09:39,880 --> 00:09:43,440
cloud native environment first 
architectures where the AI is 

207
00:09:43,440 --> 00:09:46,000
interacting directly with the 
actual infrastructure. 

208
00:09:46,000 --> 00:09:49,680
OK, but giving an AI the keys to
a live cloud environment sounds 

209
00:09:49,680 --> 00:09:52,360
terrifying for a startup, let 
alone a massive enterprise. 

210
00:09:52,360 --> 00:09:54,400
Oh, the security folks are 
definitely sweating. 

211
00:09:54,400 --> 00:09:56,800
I bet it feels like you were 
just giving it permission to 

212
00:09:56,800 --> 00:09:59,120
accidentally delete a production
database. 

213
00:09:59,200 --> 00:10:01,640
Like is anyone actually doing 
this at scale yet? 

214
00:10:01,680 --> 00:10:04,200
They absolutely are. 
Because our sources highlight 

215
00:10:04,200 --> 00:10:07,760
two completely different billion
dollar approaches from SpaceX 

216
00:10:07,760 --> 00:10:11,480
and Mozilla that prove 
enterprises are taking the leap.

217
00:10:11,640 --> 00:10:14,640
Yeah, and the scale of these 
deployments is staggering. 

218
00:10:14,680 --> 00:10:16,520
So let's look at the tooling 
layer first. 

219
00:10:16,520 --> 00:10:20,480
With SpaceX, they reportedly 
just signed a massive deal with 

220
00:10:20,560 --> 00:10:22,680
the AI coding assistant Cursor, 
OK? 

221
00:10:23,000 --> 00:10:26,000
And analysts are estimating the 
total value of this alignment 

222
00:10:26,000 --> 00:10:28,640
could be north of $10 billion 
overtime. 

223
00:10:28,640 --> 00:10:33,280
Wait, wait, wait, $10 billion 
for a coding assistant? 

224
00:10:33,280 --> 00:10:36,200
That's the estimate. 
SpaceX is managing Starling 

225
00:10:36,200 --> 00:10:39,440
satellites and, like, literal 
rocket trajectories. 

226
00:10:39,440 --> 00:10:43,520
Yeah, handing that level of code
complexity over to an AI seems 

227
00:10:43,520 --> 00:10:46,360
like a massive gamble. 
It is an incredibly aggressive 

228
00:10:46,360 --> 00:10:48,080
play. 
Some analysts called it a gutsy 

229
00:10:48,080 --> 00:10:49,800
Hail Mary. 
But you have to consider the 

230
00:10:49,800 --> 00:10:52,480
operational reality of SpaceX, 
which is what they have 

231
00:10:52,480 --> 00:10:54,960
thousands of engineers managing 
a code base with just 

232
00:10:55,040 --> 00:10:58,280
unprecedented complexity. 
They are betting that injecting 

233
00:10:58,360 --> 00:11:01,400
AI into the workflow is 
basically the only sustainable 

234
00:11:01,400 --> 00:11:04,320
way to manage the Starlink 
software life cycle going 

235
00:11:04,320 --> 00:11:06,520
forward. 
And Cursor isn't just auto 

236
00:11:06,520 --> 00:11:10,240
completing code anymore, it's 
indexing the entire repository. 

237
00:11:10,880 --> 00:11:14,280
But you're concerned about risk.
That's exactly why Cursor 

238
00:11:14,280 --> 00:11:17,240
simultaneously announced a deep 
partnership with Chainguard. 

239
00:11:17,240 --> 00:11:19,480
OK, Chainguard, what's their 
role here? 

240
00:11:20,040 --> 00:11:22,560
So Chainguard secures the 
software supply chain. 

241
00:11:23,120 --> 00:11:26,120
One of the biggest risks with AI
agents right now is that they 

242
00:11:26,120 --> 00:11:29,760
will just hallucinate a software
package, pull in a malicious 

243
00:11:29,760 --> 00:11:32,960
open source dependency and 
inject it right into your code. 

244
00:11:33,160 --> 00:11:35,720
Oh, that's terrifying. 
Right, so chain guards 

245
00:11:35,720 --> 00:11:39,800
integration ensures the AI can 
only pull from verified secure 

246
00:11:39,800 --> 00:11:43,120
container registries. 
It puts a hard guardrail on what

247
00:11:43,120 --> 00:11:45,920
the AI can import. 
OK, so SpaceX is basically 

248
00:11:45,920 --> 00:11:49,280
buying a tool to make their 
human engineers exponentially 

249
00:11:49,280 --> 00:11:50,960
faster while trying to keep it 
secure. 

250
00:11:51,000 --> 00:11:53,040
Exactly. 
A massive productivity play. 

251
00:11:53,240 --> 00:11:55,880
But contrast that with what 
Mozilla is doing, because they 

252
00:11:55,880 --> 00:11:57,840
are taking a completely 
different track here. 

253
00:11:57,840 --> 00:11:59,880
Yes, Mozilla's approach is. 
Fascinating. 

254
00:11:59,880 --> 00:12:03,640
Mozilla took an early version of
Claude Mythos Preview, which is 

255
00:12:03,640 --> 00:12:06,960
this highly danced variant of 
Anthropics model, and they 

256
00:12:06,960 --> 00:12:09,240
pointed it directly at the 
Firefox codebase. 

257
00:12:09,240 --> 00:12:10,680
The whole codebase. 
Right. 

258
00:12:10,760 --> 00:12:13,840
And the result? 
They shipped Firefox 150 with 

259
00:12:13,840 --> 00:12:18,280
fixes for 271 vulnerabilities 
that were identified and patched

260
00:12:18,360 --> 00:12:20,680
entirely by the AI. 
It's incredible. 

261
00:12:20,680 --> 00:12:23,720
Bobby Holly at Mozilla called it
quote a hopeful one for the 

262
00:12:23,720 --> 00:12:27,400
future. 
But I have to ask, how does an 

263
00:12:27,480 --> 00:12:32,840
AI actually find a semantic bug 
in legacy C++ code that actual 

264
00:12:32,840 --> 00:12:35,560
human auditors missed? 
Well, this is where we really 

265
00:12:35,560 --> 00:12:39,040
see the power of massive context
windows combined with that 

266
00:12:39,040 --> 00:12:40,640
agentic reasoning we were 
talking about. 

267
00:12:41,280 --> 00:12:45,360
Think about Firefox. 
It is an ancient, sprawling code

268
00:12:45,360 --> 00:12:47,840
base. 
Decades of code. 

269
00:12:48,360 --> 00:12:52,120
A human security auditor can 
only really old a tiny fraction 

270
00:12:52,120 --> 00:12:54,400
of the code logic in their head 
at one time, right? 

271
00:12:54,400 --> 00:12:55,840
You're just looking at one file 
at a time. 

272
00:12:55,840 --> 00:12:58,160
Exactly. 
A human might look at a specific

273
00:12:58,160 --> 00:13:01,800
function and verify that the 
local syntax is secure, but 

274
00:13:01,800 --> 00:13:04,760
Claude Mythos can ingest 
millions of tokens of context 

275
00:13:04,760 --> 00:13:06,000
all at once. 
Oh wow. 

276
00:13:06,000 --> 00:13:08,280
It doesn't just read the 
function, it traces the data 

277
00:13:08,280 --> 00:13:11,480
flow from the user interface 
through the networking layer all

278
00:13:11,480 --> 00:13:14,200
the way down to the memory 
allocation in C++. 

279
00:13:14,200 --> 00:13:18,320
R so it sees the whole map. 
Yes, it finds semantic gaps, 

280
00:13:18,320 --> 00:13:22,680
situations where the code in one
file is written correctly, but 

281
00:13:22,680 --> 00:13:26,080
the underlying logic allows for 
a memory leak or a race 

282
00:13:26,080 --> 00:13:28,840
condition when it interacts with
a completely different file 

283
00:13:28,840 --> 00:13:30,240
across the system. 
OK. 

284
00:13:30,240 --> 00:13:33,160
So we really have two distinct 
evolutions happening side by 

285
00:13:33,160 --> 00:13:33,960
side. 
Here we do. 

286
00:13:34,320 --> 00:13:37,640
Cursor is the ultimate 
productivity play right making 

287
00:13:37,640 --> 00:13:42,360
the human developer faster. 
But Mozilla is using the AI as 

288
00:13:42,360 --> 00:13:47,240
this untiring automated senior 
auditor that just holds the 

289
00:13:47,240 --> 00:13:49,880
entire system map in its memory.
Precisely. 

290
00:13:49,960 --> 00:13:52,800
We basically have the silicon 
intern writing the code and the 

291
00:13:52,800 --> 00:13:54,960
silicon senior developer 
auditing the code. 

292
00:13:55,000 --> 00:13:57,920
And that dynamic that leads 
directly to the most complex 

293
00:13:57,920 --> 00:14:00,480
engineering challenge of this 
entire shift, which is what 

294
00:14:00,480 --> 00:14:03,960
we've established that AI agents
are now operating in live cloud 

295
00:14:03,960 --> 00:14:05,240
environments. 
They're deploying 

296
00:14:05,240 --> 00:14:08,120
infrastructure, they're patching
massive code bases. 

297
00:14:08,560 --> 00:14:12,160
But modern enterprise software 
isn't built by one agent doing 1

298
00:14:12,160 --> 00:14:14,160
task. 
What what happens when you have 

299
00:14:14,160 --> 00:14:16,800
multiple specialized agents 
running around a cloud 

300
00:14:16,800 --> 00:14:19,240
environment doing completely 
different things at the exact 

301
00:14:19,240 --> 00:14:21,280
same time? 
It sounds like absolute chaos. 

302
00:14:21,280 --> 00:14:23,560
It is. 
And the Intuit engineers confirm

303
00:14:23,560 --> 00:14:25,400
it. 
There's a recent Stack Overflow 

304
00:14:25,400 --> 00:14:29,200
podcast featuring Chase Rusin 
and Steven Kalisa who run 

305
00:14:29,200 --> 00:14:34,480
engineering over at Intuit and 
they flat out called Multi agent

306
00:14:34,480 --> 00:14:36,920
orchestration. 
Quote the hardest problem in 

307
00:14:36,920 --> 00:14:38,400
engineering right now and you 
have. 

308
00:14:38,400 --> 00:14:41,840
To remember, Intuit operates 
massively complex, highly 

309
00:14:41,840 --> 00:14:44,520
regulated financial systems. 
Taxes and accounting, yeah. 

310
00:14:44,720 --> 00:14:48,080
So if you have an orchestration 
layer managing multiple agents, 

311
00:14:48,080 --> 00:14:50,520
the friction points just 
compound instantly. 

312
00:14:51,080 --> 00:14:54,880
How do you route a complex task 
to the right specialized agent? 

313
00:14:55,120 --> 00:14:58,040
And more importantly, how do you
manage domain interference? 

314
00:14:58,040 --> 00:14:59,560
Domain interference. 
Explain that for a second. 

315
00:15:00,160 --> 00:15:03,560
OK, let's say agent A is tasked 
with optimizing a database 

316
00:15:03,560 --> 00:15:07,600
schema to make it run faster. 
OK, simultaneously agent B is 

317
00:15:07,600 --> 00:15:10,640
tasked with updating the API 
endpoint that reads from that 

318
00:15:10,640 --> 00:15:13,240
exact same database. 
Oh, I see where this is going. 

319
00:15:13,480 --> 00:15:16,760
If they do not share a perfect 
context window, Agent A will 

320
00:15:16,760 --> 00:15:20,320
change the database columns and 
Agent D will deploy an API that 

321
00:15:20,320 --> 00:15:22,800
is suddenly looking for columns 
that no longer exists. 

322
00:15:22,800 --> 00:15:25,240
And everything crashes. 
They break the production 

323
00:15:25,240 --> 00:15:29,120
environment and neither agent 
realizes they did anything wrong

324
00:15:29,320 --> 00:15:31,400
because they were both just 
executing their specific 

325
00:15:31,400 --> 00:15:33,320
prompts. 
I love this analogy because it 

326
00:15:33,320 --> 00:15:35,720
really grounds the abstract 
technology for you. 

327
00:15:36,000 --> 00:15:39,160
Like building a multi agent 
system right now is essentially 

328
00:15:39,160 --> 00:15:43,280
like putting 50 brilliant, 
highly caffeinated interns in A 

329
00:15:43,280 --> 00:15:44,760
room. 
Yes, exactly. 

330
00:15:44,760 --> 00:15:47,360
You give them all corporate 
credit cards and root access to 

331
00:15:47,360 --> 00:15:50,120
the live servers, but you 
completely forget to hire a 

332
00:15:50,120 --> 00:15:52,480
manager to actually coordinate 
their projects. 

333
00:15:52,480 --> 00:15:54,360
They're just going to step all 
over each other. 

334
00:15:54,360 --> 00:15:57,320
Exactly. 
Which brings us to the protocol 

335
00:15:57,320 --> 00:15:59,120
that is desperately trying to be
that manager. 

336
00:15:59,640 --> 00:16:04,120
ISMCP. 
MCP, the Model Context protocol.

337
00:16:04,720 --> 00:16:07,400
According to that Microsoft 
developable blog post from April

338
00:16:07,400 --> 00:16:10,960
22nd, MCP is rapidly becoming 
the standardized interface for 

339
00:16:10,960 --> 00:16:13,440
all these agents. 
And MCP is so critical to 

340
00:16:13,440 --> 00:16:16,160
understand right now. 
Historically, if you wanted an 

341
00:16:16,240 --> 00:16:21,040
AI agent to say, read a GitHub 
repository, query a PostgreSQL 

342
00:16:21,040 --> 00:16:23,640
database, and then send a Slack 
message about it. 

343
00:16:23,640 --> 00:16:26,280
You had to write custom API 
integrations for all three of 

344
00:16:26,280 --> 00:16:28,000
those tools. 
Exactly. 

345
00:16:28,240 --> 00:16:32,080
Which is a nightmare to 
maintain, but MCP acts as this 

346
00:16:32,080 --> 00:16:34,480
universal client server 
handshake OK. 

347
00:16:34,600 --> 00:16:38,000
It provides A standardized way 
for an AI to automatically 

348
00:16:38,000 --> 00:16:41,120
discover what tools are 
available in its environment and

349
00:16:41,120 --> 00:16:44,400
how to communicate with them 
without needing bespoke custom 

350
00:16:44,400 --> 00:16:48,240
code for every single service. 
So to bring our intern analogy 

351
00:16:48,240 --> 00:16:52,240
back, MCP is essentially the 
standardized ID badge system. 

352
00:16:52,320 --> 00:16:54,840
That's a good way to put it. 
It lets the building security 

353
00:16:54,840 --> 00:16:58,160
quickly verify which of the 50 
interns is allowed to enter the 

354
00:16:58,160 --> 00:17:01,520
server room and which intern is 
only allowed to access the 

355
00:17:01,520 --> 00:17:03,600
cafeteria. 
Right, it standardizes the 

356
00:17:03,600 --> 00:17:05,640
access. 
And major players are really 

357
00:17:05,640 --> 00:17:09,640
throwing their weight behind it.
Like Luca Chang from AWS was 

358
00:17:09,640 --> 00:17:13,440
just at the MCP Summit in New 
York City detailing how AWS is 

359
00:17:13,440 --> 00:17:14,839
aggressively shaping the 
standard. 

360
00:17:14,839 --> 00:17:17,800
They are pushing it hard. 
But, and there is a huge but, 

361
00:17:17,800 --> 00:17:21,119
here, the Microsoft blog post 
highlighted a massive red flag. 

362
00:17:21,200 --> 00:17:25,240
The post was literally titled 
Securing MCPA Control Plane for 

363
00:17:25,240 --> 00:17:27,839
Agent Tool Execution. 
And this is exactly where this 

364
00:17:27,839 --> 00:17:30,640
rapid adoption becomes a massive
liability. 

365
00:17:30,920 --> 00:17:33,400
Well, MCP is brilliant for 
standardization, right? 

366
00:17:34,080 --> 00:17:37,720
But by standardizing how agents 
talk to our deepest, most 

367
00:17:37,720 --> 00:17:41,560
sensitive infrastructure, we are
essentially creating a massive 

368
00:17:41,560 --> 00:17:46,120
unified attack surface. 
Think about it, if every single 

369
00:17:46,120 --> 00:17:49,240
autonomous agent uses the exact 
same protocol to query the 

370
00:17:49,240 --> 00:17:53,760
customer database, a malicious 
actor only needs to find one 

371
00:17:53,760 --> 00:17:57,840
single exploit in that protocol 
to gain systemic access across 

372
00:17:57,840 --> 00:18:00,160
the whole network. 
So if you hack the ID badge 

373
00:18:00,160 --> 00:18:02,400
printer, yeah, the keys to the 
entire building. 

374
00:18:02,480 --> 00:18:05,960
That is exactly the risk. 
Enterprise adoption of MCP is 

375
00:18:05,960 --> 00:18:08,240
exploding because the 
productivity gains of having 

376
00:18:08,240 --> 00:18:12,120
multiple agents collaborating 
are just too massive to ignore. 

377
00:18:12,120 --> 00:18:13,800
You can't put the genie back in 
the bottle, you. 

378
00:18:13,800 --> 00:18:15,760
Can't. 
But the security frameworks, the

379
00:18:15,760 --> 00:18:18,920
granular access controls, the 
actual audit logging of what 

380
00:18:18,920 --> 00:18:21,560
these agents are doing in real 
time are dangerous, lagging 

381
00:18:21,560 --> 00:18:23,600
behind. 
We're giving these agents 

382
00:18:23,600 --> 00:18:26,600
execution environments just as 
Boris Cherney recommended, but 

383
00:18:26,600 --> 00:18:29,600
the nervous system connecting 
them all is largely unmonitored.

384
00:18:29,880 --> 00:18:32,560
It is literally the ultimate 
tech industry. 

385
00:18:32,560 --> 00:18:35,200
Cliche. 
We are building the rocket ship 

386
00:18:35,200 --> 00:18:39,160
at lightspeed and just promising
to figure out the seatbelts once

387
00:18:39,160 --> 00:18:40,600
we are already in orbit. 
Pretty much. 

388
00:18:40,880 --> 00:18:43,040
But thankfully, you don't have 
to just wait and hope the 

389
00:18:43,040 --> 00:18:44,640
vendors figure out the seatbelts
for you. 

390
00:18:44,640 --> 00:18:46,640
No, you can learn it yourself. 
Exactly. 

391
00:18:47,200 --> 00:18:48,880
If you were listening to this 
right now and you want to 

392
00:18:48,880 --> 00:18:51,360
understand how the mechanics of 
orchestration actually work 

393
00:18:51,360 --> 00:18:55,400
under the hood, the Microsoft 
Developer Blog published A 

394
00:18:55,440 --> 00:18:58,280
phenomenal, completely free 
resource. 

395
00:18:58,400 --> 00:19:01,640
Oh, the Lang chain course. 
Yes, it's an open source course 

396
00:19:01,640 --> 00:19:04,000
called Lang chain dot JS for 
beginners. 

397
00:19:04,800 --> 00:19:07,960
It is 8 chapters long, it 
features over 70 runnable 

398
00:19:07,960 --> 00:19:11,600
TypeScript examples, and it 
dives deeply into how to 

399
00:19:11,600 --> 00:19:15,520
basically build the manager for 
those 50 interns. 

400
00:19:15,520 --> 00:19:18,040
It's a great resource. 
It covers reasoning, tool 

401
00:19:18,040 --> 00:19:21,080
calling and how to safely 
orchestrate multiple agents. 

402
00:19:21,320 --> 00:19:23,880
So it is highly practical if you
want to get your hands dirty and

403
00:19:23,880 --> 00:19:25,960
actually build this architecture
yourself. 

404
00:19:26,120 --> 00:19:29,040
And honestly, the importance of 
a resource like that really 

405
00:19:29,040 --> 00:19:30,920
cannot be overstated right now. 
Why do you say? 

406
00:19:30,960 --> 00:19:34,280
That because you cannot secure a
system you do not understand, 

407
00:19:35,120 --> 00:19:39,240
relying purely on off the shelf 
black box agent orchestration 

408
00:19:39,520 --> 00:19:42,400
just means you are inheriting 
all of its hidden vulnerability.

409
00:19:42,400 --> 00:19:44,000
You're flying blind. 
Exactly. 

410
00:19:44,640 --> 00:19:47,320
By understanding the underlying 
architecture of link chain and 

411
00:19:47,320 --> 00:19:51,280
MCP, developers can actually 
build the necessary guardrails 

412
00:19:51,280 --> 00:19:54,080
to ensure their agents don't 
accidentally, you know, take 

413
00:19:54,080 --> 00:19:57,080
down their own infrastructure. 
OK, we have covered a massive 

414
00:19:57,080 --> 00:19:58,680
amount of ground today. 
We really have. 

415
00:19:58,840 --> 00:20:02,080
We started by diagnosing phantom
bugs and realizing that the 

416
00:20:02,080 --> 00:20:04,400
clumsy waiter the execution 
harness. 

417
00:20:04,800 --> 00:20:08,000
Was dropping the plates which 
proved why we all need to build 

418
00:20:08,000 --> 00:20:11,680
our own rigorous evil sets. 
Then we travelled through the 

419
00:20:11,680 --> 00:20:15,400
death of the local IDE, 
unpacking why AI absolutely 

420
00:20:15,400 --> 00:20:18,520
requires the real ocean of a 
live cloud environment to 

421
00:20:18,520 --> 00:20:20,720
actually observe and correct its
own mistakes. 

422
00:20:21,280 --> 00:20:23,680
We saw the sheer scale of the 
shift, with Spacex's 

423
00:20:23,680 --> 00:20:27,920
jaw-dropping $10 billion bet on 
Cursor and Mozilla proving that 

424
00:20:28,000 --> 00:20:31,840
AI can act as an untiring senior
auditor for legacy code bases. 

425
00:20:31,880 --> 00:20:34,720
Huge milestones. 
And then we ended up staring 

426
00:20:35,080 --> 00:20:39,200
right at the frontier, the 
incredibly complex, wildly 

427
00:20:39,200 --> 00:20:43,440
unsecured world of multi agent 
orchestration and MCP. 

428
00:20:43,800 --> 00:20:46,640
And, you know, the sheer 
velocity of this change really 

429
00:20:46,640 --> 00:20:50,440
forces us to confront a larger 
reality here, which is, well, if

430
00:20:50,440 --> 00:20:53,960
we are rapidly moving toward a 
near future where autonomous AI 

431
00:20:53,960 --> 00:20:57,360
agents are natively writing the 
code, provisioning the cloud 

432
00:20:57,360 --> 00:21:00,680
infrastructure, and auditing the
system for vulnerabilities all 

433
00:21:00,680 --> 00:21:04,360
on their own, what exactly is 
the future role of the human 

434
00:21:04,360 --> 00:21:05,600
software engineer? 
Oh. 

435
00:21:05,960 --> 00:21:07,800
Man right? 
Will tomorrow's developer be 

436
00:21:07,800 --> 00:21:11,160
writing loops and functions, or 
will they essentially transition

437
00:21:11,160 --> 00:21:13,040
into a silicon HR manager? 
A. 

438
00:21:13,040 --> 00:21:16,160
Silicon HR manager. 
Wow, a manager who basically 

439
00:21:16,160 --> 00:21:19,960
just spends their days defining 
agent permissions, resolving 

440
00:21:19,960 --> 00:21:23,120
conflicts between different AI 
tools, and obsessively auditing 

441
00:21:23,120 --> 00:21:24,840
their digital workforces 
performance. 

442
00:21:24,840 --> 00:21:27,360
That is, I mean, that is a 
staggering thought to leave on. 

443
00:21:27,440 --> 00:21:30,560
From writing code and a local 
text editor to essentially 

444
00:21:30,560 --> 00:21:34,080
managing an army of silicon 
employees, the the ground really

445
00:21:34,080 --> 00:21:36,120
is shifting beneath our feet. 
It really is. 

446
00:21:36,280 --> 00:21:38,320
Well, thank you so much for 
joining us on this deep dive. 

447
00:21:38,520 --> 00:21:41,400
Make sure you check out that 
Lang chain dot JS course. 

448
00:21:41,560 --> 00:21:44,120
Start experimenting with your 
own cloud native sandboxes and 

449
00:21:44,200 --> 00:21:46,880
definitely keep a close eye on 
your AI waiters. 

450
00:21:47,120 --> 00:21:48,040
We'll catch you next time.