AI SREs, Chat With Your Infrastructure with Anyshift

1
00:00:00,000 --> 00:00:04,070
I'd love to really have like morning
summaries, that are, that essentially read

2
00:00:04,075 --> 00:00:09,865
to me by either a British or Australian
accent AI that because I gotta have a

3
00:00:10,045 --> 00:00:10,985
the killer feature.

4
00:00:11,045 --> 00:00:12,555
Let's let's strip this.

5
00:00:12,785 --> 00:00:14,225
I got a French accent over here.

6
00:00:14,225 --> 00:00:16,565
I got a South African accent over here.

7
00:00:16,615 --> 00:00:20,065
I want all the robots to have more
personality so that I, I recognize their

8
00:00:20,065 --> 00:00:23,625
voice when we're in a group call with
nothing but me and the agents, and one

9
00:00:23,625 --> 00:00:26,835
of them knows my infrastructure and one
of them knows my GitHub situation and

10
00:00:27,324 --> 00:00:29,664
One of them's cooking my,
my breakfast downstairs.

11
00:00:29,664 --> 00:00:30,594
I don't know what's gonna happen.

12
00:00:33,971 --> 00:00:36,251
Welcome back to the
Agentic DevOps podcast.

13
00:00:36,251 --> 00:00:40,571
I am your host, Bret Fisher., I'm
excited to talk about anyshift.io,

14
00:00:41,391 --> 00:00:43,671
the co-founders Roxane Fischer and

15
00:00:43,671 --> 00:00:50,106
Stephane Jourdan are creating a AI SRE,
and I think I said at the beginning

16
00:00:50,106 --> 00:00:53,226
of the year, I predicted somewhere
around the beginning of the year that

17
00:00:53,226 --> 00:00:57,096
in 2025, it was the year of agents,
it was the year of Claude Code.

18
00:00:57,096 --> 00:01:02,796
It was us figuring out how to move
beyond chatbots and LSP tab completion

19
00:01:02,796 --> 00:01:07,206
of AI to actually having conversations
with AI and generating, through

20
00:01:07,206 --> 00:01:08,766
agents, generating code, right?

21
00:01:08,766 --> 00:01:12,846
And also generating YAML and
HCL and markdown and all the

22
00:01:12,846 --> 00:01:14,166
things that we do in DevOps.

23
00:01:14,496 --> 00:01:17,016
But this is the year that
we figure out context.

24
00:01:17,016 --> 00:01:19,986
And Anyshift is an
perfect example of that.

25
00:01:20,016 --> 00:01:25,536
They're a year old, roughly a year old
startup that is dealing with the context

26
00:01:25,536 --> 00:01:30,506
problem of infrastructure in that if
you want a current sense of the entire

27
00:01:30,506 --> 00:01:33,626
infrastructure, and if you're gonna
shove that into the context of an AI,

28
00:01:34,406 --> 00:01:35,816
you're gonna need to do some things.

29
00:01:35,816 --> 00:01:40,346
You're gonna need to have read only
keys to a w, your cloud, AWS whatever.

30
00:01:40,346 --> 00:01:41,786
You're gonna need to
have access to GitHub.

31
00:01:42,236 --> 00:01:44,996
You might need access to monitoring
solutions, logging solutions.

32
00:01:45,326 --> 00:01:48,146
Uh, you might need access to git
ops and deployment solutions.

33
00:01:48,146 --> 00:01:50,036
And you need to gather
all that data together.

34
00:01:50,276 --> 00:01:54,206
You need to create summaries and
memories and basically a bunch of

35
00:01:54,206 --> 00:01:58,211
tokens that you're gonna have to give
the AI to help it understand what's

36
00:01:58,211 --> 00:02:02,111
going on in infrastructure, because
AI is coming for infrastructure too.

37
00:02:02,111 --> 00:02:05,051
It's coming for DevOps and
SREs and platform engineers.

38
00:02:05,141 --> 00:02:08,621
And that's why we started this podcast
a year ago was we, we kind of predicted

39
00:02:08,621 --> 00:02:13,121
that it was coming for developers in code
first because that, that was sort of the,

40
00:02:13,121 --> 00:02:18,162
the nature of what the model companies
were providing us was they were focusing

41
00:02:18,177 --> 00:02:22,587
on code, but now people are taking that
and figuring out how do we shove the

42
00:02:22,587 --> 00:02:29,712
context of operations into the context
window of these AI LLMs and get out some

43
00:02:29,802 --> 00:02:34,538
usable data like troubleshooting, like
predicting different sorts of outages

44
00:02:34,538 --> 00:02:36,909
that might potentially happen soon.

45
00:02:36,909 --> 00:02:40,209
And like how to ensure that we're
fixing things so that errors

46
00:02:40,209 --> 00:02:41,319
and issues don't happen again.

47
00:02:41,589 --> 00:02:44,709
All the stuff that's the
concern of SRE and operators.

48
00:02:45,208 --> 00:02:49,258
So, uh, we dived in deep as basically
we talk about the little bit of the

49
00:02:49,258 --> 00:02:53,368
features and the, the reason why Anyshift
exists, but I was much more interested

50
00:02:53,368 --> 00:02:57,088
in how they're trying to solve the
problems of context management, of

51
00:02:57,088 --> 00:03:01,258
memory management, how do we avoid
hallucinations, how do we protect

52
00:03:01,258 --> 00:03:02,788
ourselves from any sort of mistakes?

53
00:03:03,088 --> 00:03:05,098
And we get deep into all
that in this episode.

54
00:03:05,098 --> 00:03:08,101
So I am glad to have them on
the show and let's get into it.

55
00:03:08,534 --> 00:03:09,434
welcome to the show.

56
00:03:09,434 --> 00:03:13,724
So on the right over there we've got
Roxane Fischer, no relation, the CEO

57
00:03:13,724 --> 00:03:16,354
and co-founder of Anyshift, anyshift.io.

58
00:03:17,474 --> 00:03:18,674
And there in the middle.

59
00:03:18,754 --> 00:03:22,214
Stephane Jourdan, I'm gonna, I'm
horrible at my French accents.

60
00:03:22,214 --> 00:03:24,654
I'm gonna try to do better
next time, who's the CTO.

61
00:03:24,654 --> 00:03:26,614
So these two co-founded Anyshift.

62
00:03:26,684 --> 00:03:27,464
welcome to the show.

63
00:03:27,964 --> 00:03:28,424
Thank for.

64
00:03:29,549 --> 00:03:29,999
Yeah.

65
00:03:30,339 --> 00:03:33,869
Roxane, so when did you all start this?

66
00:03:33,919 --> 00:03:37,044
when, how, what's the
born on date of Anyshift?

67
00:03:37,499 --> 00:03:39,089
Ah, it's a long story.

68
00:03:39,209 --> 00:03:42,579
let's say that the premises were
two years ago when I met Stephane

69
00:03:42,959 --> 00:03:44,549
in a small coffee in Paris.

70
00:03:45,049 --> 00:03:49,039
We are both at the starting point where
we wanted to create a new journey.

71
00:03:49,609 --> 00:03:55,909
Stephane with his production background
in AI and me with my AI background on a

72
00:03:55,909 --> 00:04:01,359
deep tech problem to solve, and from day
one, we knew we wanted to do something

73
00:04:01,359 --> 00:04:06,159
about data and context, even before
like the context engineering trend.

74
00:04:06,659 --> 00:04:08,279
And that's how it started.

75
00:04:08,779 --> 00:04:09,409
Nice.

76
00:04:09,649 --> 00:04:10,544
Did you say coffee shop?

77
00:04:11,044 --> 00:04:12,694
Very cute coffee shop in Paris.

78
00:04:13,879 --> 00:04:19,009
So, Stephane, tell me what's the
elevator pitch for, like, let's say

79
00:04:19,009 --> 00:04:23,239
you're a platform engineer, or SRE
type, you've got a managed production.

80
00:04:23,739 --> 00:04:26,559
You've gotta maintain
Kubernetes and the cloud stuff.

81
00:04:26,559 --> 00:04:29,499
You're luckily not someone who's
settled with this as a part-time

82
00:04:29,499 --> 00:04:30,849
job amongst many other jobs.

83
00:04:30,849 --> 00:04:33,489
Like you're a dedicated ops engineer.

84
00:04:33,749 --> 00:04:35,069
What is Anyshift gonna do for me?

85
00:04:35,569 --> 00:04:40,719
Yeah, so what I would say to this
person is that in 2026, our jobs did

86
00:04:40,719 --> 00:04:44,969
not become easier and they actually
became like much more complex.

87
00:04:45,239 --> 00:04:49,889
We are managing now, like probably
much more services, everything, the

88
00:04:49,889 --> 00:04:51,449
architectures became more complex.

89
00:04:51,449 --> 00:04:55,349
We have like many different clusters, auto
scalers, we have serverless things, we now

90
00:04:55,349 --> 00:04:57,699
have like agents taking decisions for us.

91
00:04:57,699 --> 00:05:00,129
So things became very, very complex.

92
00:05:00,439 --> 00:05:05,749
Our teams did not, become larger,
probably to the contrary actually.

93
00:05:06,229 --> 00:05:09,199
And so we really need,
like that knowledge.

94
00:05:09,199 --> 00:05:14,769
We need to offload a lot of things
to very capable agents or services.

95
00:05:14,799 --> 00:05:18,319
And luckily, we have like a set
of products that we can build,

96
00:05:18,369 --> 00:05:21,519
a lot of things around knowledge
and solving our problems and

97
00:05:22,019 --> 00:05:23,569
managing the complexity for us.

98
00:05:24,069 --> 00:05:28,234
Yeah, a year ago, I think maybe it
was like a year and a half ago, I

99
00:05:28,234 --> 00:05:32,834
started seeing companies showing up
at KubeCon or just in the operator

100
00:05:32,834 --> 00:05:36,254
space talking about managing agents.

101
00:05:36,254 --> 00:05:42,094
And I was not, I was naive and I was
new to all of it, like I, a lot of us,

102
00:05:42,094 --> 00:05:45,904
but I was like, everybody was using
this word agents and overloading it.

103
00:05:45,904 --> 00:05:48,874
And I still feel like it's used
for just talking about a dozen

104
00:05:48,874 --> 00:05:50,134
different unrelated things.

105
00:05:50,614 --> 00:05:53,814
But at the time I felt like there
these companies that were letting us

106
00:05:53,814 --> 00:05:56,814
host agents that we were gonna run in
our infrastructure, run in production.

107
00:05:57,314 --> 00:06:00,494
And maybe not managing production
with it, but just agents running

108
00:06:00,494 --> 00:06:03,404
somewhere in, in our infrastructure.

109
00:06:03,494 --> 00:06:08,934
And then around the same time, probably
even up to three years ago, at KubeCon,

110
00:06:08,934 --> 00:06:13,424
we started hearing people talking
about, even in the keynotes, I felt

111
00:06:13,424 --> 00:06:16,094
like we were joking that the keynotes
were all about AI when no one was

112
00:06:16,094 --> 00:06:18,134
actually using a or running AI yet.

113
00:06:18,404 --> 00:06:22,594
But they were really focused on the
inference side and the model building

114
00:06:22,594 --> 00:06:29,504
and managing GPUs and the sort of running
of AI infrastructure, which I didn't

115
00:06:29,504 --> 00:06:31,334
have, I didn't have a dog in that hunt.

116
00:06:31,334 --> 00:06:33,764
Like I wasn't really doing any of that.

117
00:06:33,764 --> 00:06:37,574
None of my customers or clients were
managing their own GPU infrastructure.

118
00:06:37,574 --> 00:06:41,034
They were just sort of using the
SOTA models out there and all of the

119
00:06:41,034 --> 00:06:42,444
APIs that everyone else provided.

120
00:06:42,944 --> 00:06:47,654
And so it, I felt like it wasn't really
until about a year ago that we started,

121
00:06:47,654 --> 00:06:52,404
I started to see startups at the show
that were saying, we're not here to

122
00:06:52,404 --> 00:06:55,514
help you manage the AI infrastructure.

123
00:06:55,514 --> 00:06:59,804
We're here to help you use AI
to manage the infrastructure.

124
00:06:59,804 --> 00:07:03,844
Which is a subtle difference of words,
but completely separate jobs, right?

125
00:07:04,144 --> 00:07:08,814
What gave you, what was the
initial like, obviously it's hard.

126
00:07:09,264 --> 00:07:11,964
I feel like this is a hard problem
to solve, otherwise we wouldn't have

127
00:07:11,964 --> 00:07:13,584
entire startups dedicated to it.

128
00:07:13,984 --> 00:07:19,234
What is one of the biggest challenges
to using besides just running your

129
00:07:19,234 --> 00:07:23,134
AI and your local harness and having
it fill out your HCL and YAML files

130
00:07:23,134 --> 00:07:26,274
and, you know, executing CLI for you?

131
00:07:26,514 --> 00:07:30,834
Like what's the hard part about it
seeing my infrastructure and making

132
00:07:30,834 --> 00:07:32,304
intelligent decisions about that?

133
00:07:32,976 --> 00:07:34,626
I really believe it'll be the context.

134
00:07:34,806 --> 00:07:39,866
It took us one year, even a little bit
more than that to build this context

135
00:07:39,866 --> 00:07:44,996
graph that is underneath Anyshift,
which is a reconciliation between

136
00:07:44,996 --> 00:07:47,096
dozens of different sources of data.

137
00:07:47,596 --> 00:07:51,796
Today, where we really specialize
would be how do we make one

138
00:07:51,796 --> 00:07:57,506
unified source of truth, full
time and history, of your context.

139
00:07:57,936 --> 00:08:02,796
Context means cloud providers, Kubernetes,
clusters cut basis infrastructure

140
00:08:02,796 --> 00:08:05,046
as cloud basis, monitoring host.

141
00:08:05,496 --> 00:08:10,446
And when you do, you need to do day-to-day
task, related to your production,

142
00:08:10,836 --> 00:08:15,846
you need to have the understanding of
which configuration had an impact and

143
00:08:15,846 --> 00:08:21,746
a blast radius impact on the service
or like on this specific error.

144
00:08:22,886 --> 00:08:24,686
Everything that happened
full time and needs to be.

145
00:08:24,866 --> 00:08:30,656
And so how do you manage to get these
dependencies, this connection full time?

146
00:08:31,156 --> 00:08:36,506
Today, there are a lot of like AI ops
solutions, a lot of them are plugged

147
00:08:36,536 --> 00:08:37,946
through different sources of data.

148
00:08:38,446 --> 00:08:44,596
What was really hard for us, yeah, you can
see this graph that we're building, is how

149
00:08:44,596 --> 00:08:48,676
do you make the connection between those
different universes, we call it universe.

150
00:08:49,156 --> 00:08:53,246
here you can see AWS GitHub
Kubernetes, to understand when you

151
00:08:53,246 --> 00:08:55,016
make a change, what the blast radius.

152
00:08:55,446 --> 00:08:59,296
When you need to deal with production
at scale, this context is booming.

153
00:08:59,356 --> 00:09:02,446
It's like 10 of billion of context token.

154
00:09:02,896 --> 00:09:05,536
You cannot put that in one single LLM.

155
00:09:06,036 --> 00:09:10,586
And also if you would put all this data,
like imagine like you could put all this

156
00:09:10,586 --> 00:09:15,551
data in one context window, you would only
get the correlation between data sources.

157
00:09:16,051 --> 00:09:20,691
But when you need, you need to deal
with incidents or like make a change,

158
00:09:20,841 --> 00:09:21,861
you don't need the correlation.

159
00:09:21,891 --> 00:09:22,821
You need the causality.

160
00:09:23,321 --> 00:09:28,221
And building this causality
graph is what was really tough,

161
00:09:28,371 --> 00:09:29,421
and to get some time to build.

162
00:09:29,921 --> 00:09:34,301
Yeah, I feel like right now we've
got developers very much focused on

163
00:09:34,301 --> 00:09:39,521
the local harness and we, I just, in
our Agentic DevOps Guild call this

164
00:09:39,521 --> 00:09:44,011
week, we were talking about, like
what goes in the repo now, right?

165
00:09:44,011 --> 00:09:47,401
It seems like we're like, one of the
early answers that we have is just dump a

166
00:09:47,401 --> 00:09:53,601
bunch of markdown of every documentation,
plans, intent, schemas, specs, like just,

167
00:09:53,601 --> 00:09:55,701
it's all going in the repo with the code.

168
00:09:56,121 --> 00:10:02,701
And I, I know that that's not really
scalable for those of us on the ops side.

169
00:10:02,701 --> 00:10:05,761
Like, we have things everywhere
and multiple systems.

170
00:10:05,761 --> 00:10:07,861
MCP isn't gonna solve
all those issues for us.

171
00:10:07,861 --> 00:10:11,331
And we don't all have, like,
like there isn't a 10 million

172
00:10:11,331 --> 00:10:13,741
token AI model that I'm aware of.

173
00:10:13,741 --> 00:10:19,151
Like we, we don't have infinite abilities
to just throw everything at an AI prompt

174
00:10:19,151 --> 00:10:21,041
and say, okay, now figure all this out.

175
00:10:21,041 --> 00:10:21,371
Right?

176
00:10:21,371 --> 00:10:26,741
And so I think the problem isn't
as well defined for those of us.

177
00:10:26,741 --> 00:10:31,001
Like, there's not as much people besides
as many people talking about solutions and

178
00:10:31,001 --> 00:10:32,561
architectures on how they've solved this.

179
00:10:32,561 --> 00:10:34,991
I expect that to probably be solved in
the next, like, next year where we're

180
00:10:34,991 --> 00:10:38,801
gonna see a, like a ramp up of a lot
more content of people telling stories.

181
00:10:38,801 --> 00:10:42,431
But even right now, I feel like
there's pretty, there's not a lot of

182
00:10:42,431 --> 00:10:47,671
stories out there of people talking
about how exactly they used AI to

183
00:10:47,731 --> 00:10:50,671
automate their infrastructure or
to accelerate their troubleshooting

184
00:10:50,671 --> 00:10:56,216
or to, you know, and reduce the
number of pager duty midnight calls.

185
00:10:56,276 --> 00:10:56,486
Right.

186
00:10:56,636 --> 00:10:57,596
Like that kind of thing.

187
00:10:57,956 --> 00:11:02,866
And I'm curious, like if you've already
got customers, like what's the, what

188
00:11:02,866 --> 00:11:05,686
are some of the, the callouts, what are
some of the big stories that you can

189
00:11:05,686 --> 00:11:10,126
talk about in terms of how this helped
teams or gave them like an aha moment?

190
00:11:10,706 --> 00:11:15,326
that's definitely, our customers, have
different value in the different features.

191
00:11:15,566 --> 00:11:21,076
One of the, one of the value is in
gathering all this, complex data

192
00:11:21,176 --> 00:11:23,486
very quickly, like in, in seconds.

193
00:11:23,676 --> 00:11:29,276
There's this pager duty alert ringing
and sure, as a human, as a trained

194
00:11:29,276 --> 00:11:33,696
human, you are, I dunno, you are a
senior SRE and you know all the things.

195
00:11:33,696 --> 00:11:37,806
You've been there on the job for the past
six years, for example, and great for you.

196
00:11:37,906 --> 00:11:39,406
Everything is easy for you.

197
00:11:39,406 --> 00:11:43,146
And in, I dunno, in 15 to 20
minutes, you can query the logs,

198
00:11:43,146 --> 00:11:45,126
look at the Datadog dashboards.

199
00:11:45,366 --> 00:11:48,996
You, you, you can query, you know,
who created this pool request that

200
00:11:48,996 --> 00:11:50,706
changed this specific thing, et cetera.

201
00:11:50,976 --> 00:11:51,576
Guess what?

202
00:11:51,636 --> 00:11:52,926
It's not the reality out there.

203
00:11:53,186 --> 00:11:55,646
We, we don't, we are not
always like senior engineers.

204
00:11:55,646 --> 00:11:56,786
We don't know everything.

205
00:11:57,056 --> 00:12:01,606
We probably have like this specific
deployment late at night on a Friday

206
00:12:01,606 --> 00:12:03,096
evening, and nobody knew about it.

207
00:12:03,426 --> 00:12:08,936
And having all this complexity brought
to you automatically by knowing the

208
00:12:08,936 --> 00:12:13,816
changes, knowing the deployments,
knowing the past errors that led to,

209
00:12:13,866 --> 00:12:16,936
maybe a, a causal chain of problems.

210
00:12:17,096 --> 00:12:22,976
The small, the, the slowly increasing
rate of issues that, that was driving in

211
00:12:22,976 --> 00:12:28,516
the end to another issue that actually,
I dunno, created this database downtime

212
00:12:28,516 --> 00:12:33,556
that in the end was the actual issue
for your 500 errors and on the API.

213
00:12:34,056 --> 00:12:39,986
And that's all of this is so complex
today that having a system that really

214
00:12:39,986 --> 00:12:44,506
gather all this information, for you and
present it to you as a, as something,

215
00:12:44,566 --> 00:12:48,226
that you can digest and they can,
you can really process as a human and

216
00:12:48,376 --> 00:12:51,896
you can just act upon it and resolve
that, that problem very quickly.

217
00:12:52,336 --> 00:12:52,971
Yeah, the complexity.

218
00:12:53,471 --> 00:12:57,621
So, how am I interfacing?

219
00:12:57,921 --> 00:13:01,011
This is a question I've been thinking
about a friend of mine, Viktor Farcic,

220
00:13:01,661 --> 00:13:03,401
who runs the DevOps Toolkit channel.

221
00:13:03,821 --> 00:13:07,121
And we were talking a couple weeks
ago, and he has this theory that

222
00:13:07,171 --> 00:13:11,061
I think I still agree with, that
we're gonna be so used to talking

223
00:13:11,061 --> 00:13:12,651
with the chat bot in our harness.

224
00:13:12,681 --> 00:13:17,101
Like we're essentially writing our
writing e everything, whether it's code

225
00:13:17,101 --> 00:13:19,681
or YAML or you know, documentation.

226
00:13:19,681 --> 00:13:22,951
We're writing all that through our
local harness and that that needs to

227
00:13:22,951 --> 00:13:28,071
become the primary window to everything,
rather than us having 20 different

228
00:13:28,101 --> 00:13:32,991
chat bots in 20 different siloed
places where none of like, one doesn't

229
00:13:32,991 --> 00:13:36,351
know about the other and there's no
shared context or shared memories.

230
00:13:36,871 --> 00:13:40,411
Do you see that as like the future
of like, is this plugged, is this

231
00:13:40,441 --> 00:13:44,551
something where I can plug it in to
my local harness so it can interface?

232
00:13:44,551 --> 00:13:48,211
Or am I using this like through Slack
or, I'm just like kinda thinking

233
00:13:48,301 --> 00:13:52,251
what are the different ways that I
can interface with this system as a,

234
00:13:52,521 --> 00:13:54,231
the human to AI interface, I guess?

235
00:13:54,231 --> 00:13:54,591
Yeah.

236
00:13:55,091 --> 00:13:55,381
Yeah.

237
00:13:55,456 --> 00:13:58,546
I'm smiling a lot, so let me
take this one because like I, I

238
00:13:58,546 --> 00:13:59,686
have the conversation very often.

239
00:14:00,294 --> 00:14:04,154
today you can use, our like
Anyshift for various forms.

240
00:14:04,544 --> 00:14:08,414
We have a CLI, we have an MCP,
a web app, and a Slack app.

241
00:14:08,914 --> 00:14:10,474
And it really depends on your needs.

242
00:14:10,954 --> 00:14:14,914
The web app, you can plot some
diagrams about your infrastructure.

243
00:14:15,044 --> 00:14:19,714
You can create reports about
anomalies and proactively go

244
00:14:19,714 --> 00:14:21,064
fetch from information for you.

245
00:14:21,564 --> 00:14:27,184
You can view the root cause
analysis being performed live

246
00:14:27,214 --> 00:14:31,024
for live graph of hypothesis with
all the fact that we gathered.

247
00:14:31,524 --> 00:14:32,664
You are a visual person.

248
00:14:33,164 --> 00:14:39,284
But I need to say that I guess our poor
users really love us in the CLI and MCP.

249
00:14:39,784 --> 00:14:42,504
People love Claude Code
and we love it too.

250
00:14:42,714 --> 00:14:46,674
And so like how do we bring this
context into your day-to-day workflow?

251
00:14:47,174 --> 00:14:50,654
As you were saying, what's kind of
the future like, are we still like

252
00:14:50,654 --> 00:14:53,224
in a person to agent interface?

253
00:14:53,344 --> 00:14:57,784
So like, I am a person who works in,
and I'm going to ask in the chat bot

254
00:14:57,814 --> 00:15:02,404
on one specific product, on a web app
a question, or do I want my agent to

255
00:15:02,404 --> 00:15:06,784
actually interface with another agent
to get this context to perform a task?

256
00:15:07,174 --> 00:15:10,454
And this is what we see at the
moment and how things evolve.

257
00:15:10,954 --> 00:15:16,134
We have some customers who asked us to
have the agent to agent protocol, like how

258
00:15:16,134 --> 00:15:21,884
does any, so our agents who has the entire
context about the different deployments,

259
00:15:21,934 --> 00:15:24,434
commits, all the changes full time?

260
00:15:24,764 --> 00:15:28,634
So this agent that has a context
to decide how to trigger the

261
00:15:28,634 --> 00:15:33,854
Dynatrace, agent, for instance,
to be the master of other agents.

262
00:15:34,304 --> 00:15:38,644
I don't know how fast this transition
from like completely automated agent

263
00:15:38,644 --> 00:15:43,249
to agent will happen, but there,
there are some automat automation

264
00:15:43,249 --> 00:15:47,469
happening already, and so we need to
be at different places depending of

265
00:15:47,469 --> 00:15:52,939
different needs to provide this context
to the different like agents arguments.

266
00:15:53,439 --> 00:15:55,629
So yeah, it's like meet you where you are.

267
00:15:55,629 --> 00:15:58,239
If you're used to using the
website dashboard, if that's

268
00:15:58,239 --> 00:16:00,149
your thing, you can do that.

269
00:16:00,149 --> 00:16:05,359
If you're someone who wants to use a
CLI locally, I guess I, I would maybe

270
00:16:05,359 --> 00:16:09,949
like, I'd create a skill or something
that would know about the CLI, so that

271
00:16:09,949 --> 00:16:13,129
I could just tell my local harness, Hey,
go find this information on Anyshift.

272
00:16:13,249 --> 00:16:16,639
Is that kind of the, the
idea there with the, the CLI?

273
00:16:17,139 --> 00:16:17,629
Exactly.

274
00:16:18,129 --> 00:16:18,549
Yeah.

275
00:16:19,049 --> 00:16:19,259
Yeah.

276
00:16:19,289 --> 00:16:23,639
'cause that, that maybe we haven't
stated that on the show before, but

277
00:16:24,139 --> 00:16:29,734
I feel like in 2026, a lot of what we
were trying to use MCP for a year ago,

278
00:16:30,234 --> 00:16:32,814
you know, when it was the new hotness
and everybody was MCP-ing everything

279
00:16:32,814 --> 00:16:34,014
and we had way too many MCP tools.

280
00:16:34,044 --> 00:16:37,464
I think at some point I looked at my
copilot inside or I looked at my VS

281
00:16:37,464 --> 00:16:42,264
Code, which all the extensions are
still now shipping with MCP tools and

282
00:16:42,274 --> 00:16:45,124
I was tinkering around with Copilot and
I just clicked on the MCP tools and I

283
00:16:45,124 --> 00:16:50,104
think I had like 150 tools loaded in my
context, and of which none I actually

284
00:16:50,104 --> 00:16:54,964
put in myself or knew about because they
all came with the VS Code extensions.

285
00:16:55,294 --> 00:16:58,474
And so then I had to like basically
uncheck them all because I, I didn't

286
00:16:58,474 --> 00:17:00,874
realize they were all filling up
my context window essentially.

287
00:17:01,174 --> 00:17:05,494
And now it feels like we're leaning
towards, you know, actually the

288
00:17:05,494 --> 00:17:08,494
AI's really good at just having bash
like, just give it bash, tell it

289
00:17:08,524 --> 00:17:12,994
where some tools are maybe CLI tools
and that's the way it can access.

290
00:17:12,994 --> 00:17:14,674
So I've completely shifted my workflow.

291
00:17:14,674 --> 00:17:17,964
I don't know if that's what everybody's
doing, but a lot of things that I was

292
00:17:17,964 --> 00:17:22,974
trying to use MCP for before, I'm now
just f downloading the CLI tool and

293
00:17:22,974 --> 00:17:26,094
making sure that the AI knows that it
can use it when I'm asking it a question

294
00:17:26,094 --> 00:17:29,034
rather than expecting it to use some MCP.

295
00:17:29,484 --> 00:17:33,564
It, we just recently had, I think
Google launched their Workspace CLI,

296
00:17:33,564 --> 00:17:38,184
so now I can have it access my email
and my Google Docs all through a CLI.

297
00:17:38,184 --> 00:17:43,234
And that kind of for me now becomes
the default, first thing I do is if I'm

298
00:17:43,234 --> 00:17:46,324
looking at a tool, if I'm signing up
for something like Anyshift, whether

299
00:17:46,324 --> 00:17:50,884
it's AI or not, I'm looking for them to
have a CLI, so I, I can give that to my

300
00:17:50,884 --> 00:17:54,894
harness to, to do the work rather than
trying to find out if they have an MCP.

301
00:17:54,904 --> 00:17:58,384
Do you see that same trend with your
customers and you're part of the industry?

302
00:17:58,884 --> 00:17:59,694
Yeah, yeah.

303
00:17:59,694 --> 00:18:02,629
One, one of them, I have in
mind one, one specific customer.

304
00:18:02,729 --> 00:18:07,159
He's supercharged his workflow,
he has so much to do, many, many

305
00:18:07,159 --> 00:18:08,719
different products to manage.

306
00:18:08,959 --> 00:18:12,959
So many problems to tackle
with such a small team.

307
00:18:13,139 --> 00:18:17,219
So he supercharged himself with all
the different CLIs he could use.

308
00:18:17,249 --> 00:18:21,189
He created skills very dedicated
to his way of working, et cetera.

309
00:18:21,189 --> 00:18:25,469
And it's gathering every single
bit of information of context that

310
00:18:25,469 --> 00:18:30,259
he can bring to his workstation
through his Claude Code, so he can

311
00:18:30,259 --> 00:18:34,549
like, tackle more, pro more, yeah,
problems, project products, et cetera.

312
00:18:34,759 --> 00:18:35,959
So that's really something.

313
00:18:35,959 --> 00:18:36,859
One, one case.

314
00:18:36,939 --> 00:18:41,999
We, we have, so it's, it's huge, for
him on the laptop and another use case,

315
00:18:42,029 --> 00:18:46,249
we, we actually learned, recently from
one of our largest customers as well.

316
00:18:46,299 --> 00:18:51,079
they actually use the MCP to build
something around the Anyshift MCP

317
00:18:51,079 --> 00:18:55,169
so that, that knowledge, all that
information they use it to, to create

318
00:18:55,169 --> 00:18:56,669
something very specific for them.

319
00:18:56,669 --> 00:19:01,219
So like it's super agent based on the,
that source of information and it's

320
00:19:01,219 --> 00:19:04,849
central to help them fix problems earlier.

321
00:19:04,879 --> 00:19:09,759
Because in the end, when you are managing
product and systems, you really need to

322
00:19:09,759 --> 00:19:11,589
solve the problems as fast as possible.

323
00:19:11,589 --> 00:19:15,429
And today it's, it's possible if you have
the right information at the right moment.

324
00:19:15,849 --> 00:19:20,499
And just querying logs right now,
it's probably not gonna help you.

325
00:19:20,589 --> 00:19:22,719
It's probably something that
happened like five years ago, or

326
00:19:22,719 --> 00:19:24,399
five days ago, or five hours ago.

327
00:19:24,549 --> 00:19:25,599
That was the culprit.

328
00:19:26,099 --> 00:19:27,749
So let me get, let me
make sure I understand.

329
00:19:27,749 --> 00:19:31,619
So you're saying they built their
own custom agent running somewhere

330
00:19:31,619 --> 00:19:34,829
and it has access to your MCP,
is that what you're saying?

331
00:19:35,729 --> 00:19:42,039
And so that agent is pulling data
out on its own from your graph.

332
00:19:42,399 --> 00:19:45,249
Yeah, and I think I read on your
website that your graph, like you're

333
00:19:45,249 --> 00:19:46,599
doing this all read only, right?

334
00:19:46,599 --> 00:19:48,759
Like these are, these have
read only credentials.

335
00:19:48,859 --> 00:19:51,649
We're not really talking about this thing
going and tearing down my infrastructure

336
00:19:51,649 --> 00:19:52,939
automatically for me, right?

337
00:19:52,939 --> 00:19:57,759
Like, this is something that's really
meant for investigation and analysis

338
00:19:57,759 --> 00:19:59,709
rather than command and control.

339
00:19:59,709 --> 00:20:02,169
Is that the goal here
with Anyshift this just

340
00:20:02,244 --> 00:20:03,264
well, well, well,

341
00:20:03,369 --> 00:20:03,789
taking?

342
00:20:04,289 --> 00:20:05,309
Well, let's roadmap.

343
00:20:05,359 --> 00:20:07,699
actually, we were probably too cautious.

344
00:20:07,749 --> 00:20:12,189
I was probably the, the most cautious,
between Roxane and, and I, and I was

345
00:20:12,189 --> 00:20:17,289
like, no, you can't, you, you can't do
like, destructive things so, so early.

346
00:20:17,709 --> 00:20:23,449
And actually, our customers are, asking
for it now, as, as really like they want,

347
00:20:23,549 --> 00:20:27,149
they ask for like, okay, we have this
specific runbook, we have this issue.

348
00:20:27,269 --> 00:20:32,309
It's waking up like every other Tuesday
at 3:00 AM and we know that when this

349
00:20:32,309 --> 00:20:36,239
rings, we have to do this action, and we
have no way of knowing when it's gonna

350
00:20:36,239 --> 00:20:37,619
ring, but it's gonna ring at 2:00 AM.

351
00:20:38,019 --> 00:20:42,989
If Any could just do this specific
run book, this action right now, it

352
00:20:42,989 --> 00:20:45,299
would save one guy's night's sleep.

353
00:20:45,659 --> 00:20:50,909
And so we are actually working on
making Any do actions, because people

354
00:20:50,909 --> 00:20:56,659
are actually ready, like it's April
2026 and people are ready for having

355
00:20:56,929 --> 00:21:00,619
agents do actions on their production
systems under control, obviously.

356
00:21:00,979 --> 00:21:05,499
But it was the same when you
hired someone new in your team.

357
00:21:05,499 --> 00:21:10,354
Like, you want them to be ready as fast as
possible and with all that information and

358
00:21:10,359 --> 00:21:13,349
you can do pretty accurate things already.

359
00:21:13,849 --> 00:21:17,989
Yeah, actually by the time this podcast
comes out, the Mendral one with Sam

360
00:21:17,989 --> 00:21:21,589
Alba will come out because we recorded
that, I think last week, and we had

361
00:21:21,589 --> 00:21:25,909
a similar conversation where I wanted
him, I'm using their product to help

362
00:21:25,909 --> 00:21:29,989
manage my GitHub and automate some
of the toil of GitHub management.

363
00:21:30,379 --> 00:21:33,739
And they didn't have any, right,
they didn't have any, it wasn't

364
00:21:33,739 --> 00:21:34,669
taking actions yet, right?

365
00:21:34,669 --> 00:21:37,039
It was just incident
reporting and analysis.

366
00:21:37,189 --> 00:21:39,299
And then would, give you sort of.

367
00:21:39,799 --> 00:21:42,329
The ability for it to do something,
but you had to approve, right?

368
00:21:42,329 --> 00:21:45,539
You had to let it, you had to go in and
read and make sure you wanted to do that

369
00:21:45,539 --> 00:21:46,799
thing, and then it would go do that thing.

370
00:21:46,799 --> 00:21:49,289
And the, and the thing it was doing
at most was a PR, like, it wasn't

371
00:21:49,289 --> 00:21:52,109
even, it wasn't even committing
to main or anything like that.

372
00:21:52,499 --> 00:21:53,969
And I wanted him to go faster.

373
00:21:53,969 --> 00:21:56,779
I was like, when are we gonna have the
ability for it to just auto fix, auto fix?

374
00:21:57,129 --> 00:22:00,479
And I feel like there's, like, I'm
imagining this as sort of, like a linear

375
00:22:00,479 --> 00:22:04,739
progression, a maturity model that's
really just this line that I'm dragging

376
00:22:04,739 --> 00:22:08,549
on one side, it starts in the blue,
and then as I drag it to the right,

377
00:22:08,609 --> 00:22:12,389
it's going redder and redder to like,
basically to the end is YOLO mode.

378
00:22:12,889 --> 00:22:16,979
And I feel like all of us over the last
six months have gone through that similar

379
00:22:16,979 --> 00:22:22,859
process with learning Claude Code and some
of the other ag agentic, harnesses where

380
00:22:22,989 --> 00:22:26,734
I know enough people now that I know some
people that are very much in full safety,

381
00:22:27,234 --> 00:22:30,954
they still want copilot to ask them
about every command before it does it.

382
00:22:31,404 --> 00:22:34,764
And then there's others that
are going full off on safety

383
00:22:34,764 --> 00:22:36,684
and they just let it run.

384
00:22:37,104 --> 00:22:39,954
I mean, there's obviously the
OpenClaw people that are just,

385
00:22:39,954 --> 00:22:41,304
you know, a whole nother level.

386
00:22:41,694 --> 00:22:45,879
But everyone is in their own comfort
level and I tend to find that, we

387
00:22:45,879 --> 00:22:49,839
all just like a junior engineer, we
all need to trust the model and the

388
00:22:49,839 --> 00:22:53,709
harness, and we're just gonna call
that the AI agent, and we need to trust

389
00:22:53,709 --> 00:22:55,389
it, and so we just need time with it.

390
00:22:55,389 --> 00:23:00,669
We just need to watch it not hallucinate
for months on end or weeks on end before

391
00:23:00,669 --> 00:23:01,899
we're willing to give it more work.

392
00:23:01,929 --> 00:23:04,599
And then we give it a little bit more,
and then we maybe adjust our skills

393
00:23:04,599 --> 00:23:08,039
and our Claude file or whatever we're
doing, and we get, we get a little bit

394
00:23:08,039 --> 00:23:11,009
better and we realize this thing hasn't
made a mistake in a couple of weeks

395
00:23:11,009 --> 00:23:14,159
other than just, you know, like little
innocent mistakes that a human would

396
00:23:14,159 --> 00:23:15,869
make that are not even really mistakes.

397
00:23:15,869 --> 00:23:17,609
They're just a different
choice than what I would make.

398
00:23:18,049 --> 00:23:21,819
And eventually we get to that point where
we're like, okay, I'm turning off safety.

399
00:23:21,819 --> 00:23:23,199
I don't even care about sandboxing.

400
00:23:23,199 --> 00:23:24,519
I trust this thing implicitly.

401
00:23:24,889 --> 00:23:26,149
You know, go Terraform apply.

402
00:23:26,569 --> 00:23:31,299
And, and I feel like that is, eventually,
I think, gonna happen to most or all

403
00:23:31,299 --> 00:23:34,449
of us, if at least we're allowed to
by organizations, obviously certain

404
00:23:34,449 --> 00:23:37,329
organizations are gonna put a lot of
re restraint and restriction on that.

405
00:23:37,629 --> 00:23:39,849
But I feel like we're all
on that path somewhere.

406
00:23:39,849 --> 00:23:40,899
So it's somewhere on that.

407
00:23:41,079 --> 00:23:43,089
I need to just have a graph
I put up on screen and just

408
00:23:43,089 --> 00:23:44,169
basically allow people to choose.

409
00:23:44,169 --> 00:23:44,919
Where are you?

410
00:23:45,229 --> 00:23:48,169
I just wanted to speak about trust
because like you're mentioning like how

411
00:23:48,169 --> 00:23:52,249
do you progressively trust the agent
from like, doing more and more action.

412
00:23:52,549 --> 00:23:57,259
And this is something which really believe
as well, like you need to give like human,

413
00:23:57,259 --> 00:24:01,489
and human in the loop as a capability to
understand exactly what the agent did,

414
00:24:01,579 --> 00:24:07,089
what type of queries this agent performed
to like build this progressive trust.

415
00:24:07,089 --> 00:24:11,159
And like later give the agents like
the capabilities of making actions.

416
00:24:11,659 --> 00:24:16,719
So like how do you actually like
exactly, with like a junior SAV joining

417
00:24:16,779 --> 00:24:21,269
your team, you can of understand the
progress and the type of capabilities

418
00:24:21,329 --> 00:24:26,249
this person is doing to then allow
widely selection, pull requests from

419
00:24:26,249 --> 00:24:30,299
being done and then like complete
out the loop kind of capabilities.

420
00:24:30,799 --> 00:24:31,279
Yeah.

421
00:24:31,334 --> 00:24:31,604
Yeah.

422
00:24:31,604 --> 00:24:34,844
The, to me the auditing is like,
really, it feels really important.

423
00:24:34,844 --> 00:24:39,514
Like if I'm having a, if I'm having
a robot take actions on my behalf,

424
00:24:39,544 --> 00:24:43,239
it's I have lots of things I think
I want, and I think that's all gonna

425
00:24:43,239 --> 00:24:45,889
completely change the minute I actually
start using systems, like Anyshift,

426
00:24:45,909 --> 00:24:50,909
because I, what I perceive to be what
I wanted to do, you know, I wanted to

427
00:24:50,909 --> 00:24:55,649
have like read only credentials until
I manually granted right in the moment.

428
00:24:55,649 --> 00:25:00,629
And then I wanted to swap out tokens or,
or keys so that it has the right pat now.

429
00:25:00,629 --> 00:25:03,689
And then I then kinda like we
have build and plan mode locally.

430
00:25:03,689 --> 00:25:07,499
Like I, I kind of want that
in my infrastructure where in

431
00:25:07,499 --> 00:25:10,769
certain moments I'm, for certain
activities I'm gonna give it.

432
00:25:10,769 --> 00:25:14,189
Right, but other ones, it doesn't
have that ability or doesn't even

433
00:25:14,189 --> 00:25:15,789
have access to Kubernetes API.

434
00:25:15,789 --> 00:25:16,829
Or if it does, it's read only.

435
00:25:16,829 --> 00:25:22,139
Like, I'm just sort of imagining how
I'm gonna slowly onboard this thing.

436
00:25:22,139 --> 00:25:26,609
One of the ideas I was talking with Sam
about was the idea that, if things are

437
00:25:26,609 --> 00:25:30,569
recurring or recurring types of incidents
where it's a similar type of problem

438
00:25:30,569 --> 00:25:34,689
or maybe the, the tool that it needs to
fix is something that I've, I've already

439
00:25:34,689 --> 00:25:37,989
approved for, right, or permission for
it to actually do something on, on a

440
00:25:37,989 --> 00:25:42,249
right access that it, it now knows that
there are certain things it's allowed to

441
00:25:42,249 --> 00:25:45,649
do and I've given it, we've established
those permissions that it can have.

442
00:25:45,649 --> 00:25:48,199
And then there's other things
that it's still not ready for.

443
00:25:48,199 --> 00:25:51,799
And we're, and maybe as a team we're
not ready for, it may be because

444
00:25:51,799 --> 00:25:54,349
we have crappy documentation,
so it doesn't have good context.

445
00:25:54,349 --> 00:25:58,869
It might be because we haven't given
it full read access to all that

446
00:25:58,869 --> 00:26:00,219
particular part of the infrastructure.

447
00:26:00,219 --> 00:26:03,969
Maybe it's hybrid and, you know, there's
missing components, so it tends to make

448
00:26:03,969 --> 00:26:05,439
the wrong decisions or hallucinate.

449
00:26:05,809 --> 00:26:09,539
Do you see that spectrum
happening in Anyshift?

450
00:26:09,649 --> 00:26:13,489
Is it, is it gonna be like this messy
world of like some things are this and

451
00:26:13,489 --> 00:26:16,219
some things are that, or is it just
gonna be like an on and off toggle?

452
00:26:16,219 --> 00:26:18,459
How do you see that sort of happening.

453
00:26:18,579 --> 00:26:18,969
a great question.

454
00:26:19,469 --> 00:26:23,609
Like we really see like Anyshift as a
new member that you're onboard within

455
00:26:23,609 --> 00:26:27,649
the company, you give this person
some accesses and depending of the

456
00:26:27,649 --> 00:26:33,669
accesses this person have, she or he
will be able to debug an incident or

457
00:26:33,669 --> 00:26:35,229
like to perform a day-to-day task.

458
00:26:35,729 --> 00:26:38,999
It's also like super important to
mention that when we speak about

459
00:26:38,999 --> 00:26:41,099
context, you need to have memory as well.

460
00:26:41,579 --> 00:26:45,339
So our agents, all have like
self reinforcement memory.

461
00:26:45,839 --> 00:26:48,479
They will learn from past
patterns, past incidents.

462
00:26:48,749 --> 00:26:52,829
All the data, all, everything that
they have seen very similarly to

463
00:26:53,309 --> 00:26:57,299
someone, new team that begins and
then will evolve within the team.

464
00:26:57,749 --> 00:27:02,219
it'll learn that day one, oh, I thought
like this is an incident because like

465
00:27:02,219 --> 00:27:04,349
twice a day I have huge scaling events.

466
00:27:04,799 --> 00:27:08,639
And two days later actually I understand
that this is a normal behavior because

467
00:27:08,639 --> 00:27:11,909
this is a pattern of the production
of the company I'm working at.

468
00:27:12,409 --> 00:27:15,679
That's the type of context that you
need to have and that our agents have

469
00:27:15,679 --> 00:27:21,429
access to, to be able to learn from
connection that you only find at one time.

470
00:27:21,699 --> 00:27:23,769
We spoke about the graph
that we're building.

471
00:27:23,829 --> 00:27:27,164
So this is like a time machine of
your production, all resource and

472
00:27:27,734 --> 00:27:29,654
dependencies, how things connect.

473
00:27:29,954 --> 00:27:34,364
But sometimes you don't have all this
connection, all this data from just

474
00:27:34,364 --> 00:27:38,084
like, integration and like raw hard data.

475
00:27:38,504 --> 00:27:41,054
Some of those connection can
only be found at one time.

476
00:27:41,554 --> 00:27:43,804
Like how does the
service pick another one?

477
00:27:44,224 --> 00:27:47,904
We learn it as well through
the memory at one time will

478
00:27:47,934 --> 00:27:49,584
upgrade the memory of our agents.

479
00:27:49,704 --> 00:27:50,874
They will do it by themself.

480
00:27:51,374 --> 00:27:54,314
And this also is part of
the context we're building.

481
00:27:54,814 --> 00:27:58,474
The resources, all the dependencies,
all the changes that you need to have

482
00:27:58,534 --> 00:28:04,949
between dozen of fragmented tools,
Datadog, AWS, Kubernetes, like, many ones.

483
00:28:05,399 --> 00:28:09,389
How do you bring back this context in
one unified view to debug incidents

484
00:28:09,509 --> 00:28:11,159
and also like prevent some of them?

485
00:28:11,519 --> 00:28:15,689
Prevention also like as important
than as to just react on the fire.

486
00:28:16,189 --> 00:28:20,179
I wanna ask about prevention, but before
I do, the concept of memory is something

487
00:28:20,179 --> 00:28:22,959
that I actually spent a lot of time
in the last week kind of understanding

488
00:28:22,959 --> 00:28:26,949
where the industry is going and like
even for local harnesses now, I just,

489
00:28:27,249 --> 00:28:30,389
I think it was yesterday actually, I
learned that Claude was explaining to me.

490
00:28:30,869 --> 00:28:35,189
I basically asked it list to me all the
harnesses that you can find and whether or

491
00:28:35,189 --> 00:28:39,719
not they have built-in memory components,
you know, or features inside them.

492
00:28:39,719 --> 00:28:44,219
And it was like it listed, Claude Code
has it and Copilot has it now that where,

493
00:28:44,289 --> 00:28:48,134
like Claude Code stores it, it's very,
it's a, I think it's right now, I don't

494
00:28:48,134 --> 00:28:51,974
know about team accounts, but I know
it, it stores it for you individually

495
00:28:51,974 --> 00:28:55,694
and they have this new auto dream thing
that also enhances memories over time.

496
00:28:56,174 --> 00:28:59,404
And for those of those not familiar
with this stuff, I think we're kind

497
00:28:59,404 --> 00:29:00,574
of all centering around the idea.

498
00:29:00,574 --> 00:29:05,164
We understand context, we understand
like skills and the agent files and sort

499
00:29:05,164 --> 00:29:06,814
of these different parts of the puzzle.

500
00:29:07,204 --> 00:29:11,134
But this memory, this thing we're
labeling as memory to me is about

501
00:29:11,634 --> 00:29:16,604
summarizing previous conversations
or previous sessions with an agent

502
00:29:16,664 --> 00:29:21,404
and storing those long term so
that they can be whether basically

503
00:29:21,464 --> 00:29:23,204
searched and injected dynamically.

504
00:29:23,204 --> 00:29:25,304
I think they're all kind of doing
something a little bit different

505
00:29:25,664 --> 00:29:28,934
on how exactly they bring memories
into the current conversation.

506
00:29:29,054 --> 00:29:32,304
And this is just, to me, a natural
extension of how we're going to take

507
00:29:32,304 --> 00:29:34,104
this beyond the current session.

508
00:29:34,154 --> 00:29:39,014
I've been writing a newsletter this week
around the bad habit that I, I see myself

509
00:29:39,014 --> 00:29:44,404
doing and others doing, where we start
a conversation with an agent and we just

510
00:29:44,404 --> 00:29:49,014
continue that conversation rather than
starting a new one because there's so much

511
00:29:49,014 --> 00:29:53,584
information locked up in that session that
becomes almost like an AI tribal knowledge

512
00:29:53,584 --> 00:29:55,474
at that point because it's only in there.

513
00:29:55,684 --> 00:29:56,704
My team doesn't have it.

514
00:29:56,704 --> 00:29:59,284
My other AI models
certainly don't have it.

515
00:29:59,284 --> 00:30:02,224
My other sessions don't even have
it, and it's only in that session.

516
00:30:02,674 --> 00:30:06,154
And so I, it's like, it's like a precious
little snowflake of a conversation,

517
00:30:06,559 --> 00:30:11,949
and that I need to have like almost a
personal discipline to somehow get that

518
00:30:12,009 --> 00:30:15,999
into, whether it's the agent's file or
documentation in the repo or something

519
00:30:15,999 --> 00:30:19,479
else where the agent, the future
conversations can access those memories.

520
00:30:19,749 --> 00:30:22,089
It turns out that maybe all the tools
are gonna manage this for us, and

521
00:30:22,089 --> 00:30:25,269
it's all gonna be eventually a solved
problem where maybe there'll be central

522
00:30:25,269 --> 00:30:26,829
memories for the team somewhere.

523
00:30:27,079 --> 00:30:29,659
I'm, I'm not sure how it's all gonna
shape out in terms of just the harnesses,

524
00:30:29,659 --> 00:30:31,129
but it's cool that you're doing that.

525
00:30:31,129 --> 00:30:33,649
I know that Mendral with Sam, they're
also doing the same thing where they're

526
00:30:33,649 --> 00:30:37,329
adding memories that, they're incidents
of creating memories, but also you

527
00:30:37,329 --> 00:30:40,599
can add in your own memories, which
I almost, I was telling 'em like,

528
00:30:40,599 --> 00:30:41,709
this is the tribal knowledge area.

529
00:30:41,709 --> 00:30:46,289
This is where I'm just typing in, in
my case, for him it was, I needed,

530
00:30:46,789 --> 00:30:54,289
I specifically want GitHub Actions,
reusable actions to not pin the digest

531
00:30:54,589 --> 00:30:58,189
if the, this is very technical, but
the calling action and the reusable

532
00:30:58,189 --> 00:31:02,659
action, if they're both managed by me,
I don't need to pin one to the other.

533
00:31:02,659 --> 00:31:07,129
I just need to have the reusable
action pinning any actions it's using.

534
00:31:07,129 --> 00:31:09,229
I know that's, that's if you're
not using enough actions, that

535
00:31:09,229 --> 00:31:10,159
just sounded like gibberish.

536
00:31:10,519 --> 00:31:13,459
But it was something where I have
a very specific workflow for me

537
00:31:13,459 --> 00:31:15,679
and, and the people that I work
with, and that's how we do it.

538
00:31:16,129 --> 00:31:19,519
But the linters and the sort of
LLMs think that there's another

539
00:31:19,519 --> 00:31:20,629
way that they should be doing it.

540
00:31:20,629 --> 00:31:24,289
And I need to give it very explicit
instructions that it maybe won't

541
00:31:24,289 --> 00:31:27,319
be seeing because not all these
tools are looking at a repo, right?

542
00:31:27,319 --> 00:31:31,339
They're not all harnesses like these
SRE tools and these GitOps, these,

543
00:31:31,399 --> 00:31:34,309
DevOps tools aren't necessarily
looking in repo for instructions.

544
00:31:34,729 --> 00:31:39,079
So where does it find all of
the lore of my infrastructure?

545
00:31:39,079 --> 00:31:41,749
All of the preferences that we
made over time, the architecture

546
00:31:41,749 --> 00:31:43,159
decisions that we made over time.

547
00:31:43,419 --> 00:31:46,889
Are, I'm assuming that you have a bunch
of a series of plugins or a series of

548
00:31:46,889 --> 00:31:51,059
integrations, that aren't just looking
at my, your infrastructure today, but

549
00:31:51,059 --> 00:31:55,139
are also maybe looking at like, are you
looking at Linear, at Confluence, at Jira?

550
00:31:55,139 --> 00:31:57,599
Like does it pull that kind
of information as well?

551
00:31:58,099 --> 00:31:58,939
It's even better.

552
00:31:58,969 --> 00:32:02,789
I will let Roxane be more specific
maybe about the mechanism of the

553
00:32:02,789 --> 00:32:06,239
memory itself, like she is much
more, compared than me on this one.

554
00:32:06,629 --> 00:32:08,009
but something that Anyshift does.

555
00:32:08,009 --> 00:32:08,789
So, yeah, you're right.

556
00:32:08,889 --> 00:32:13,149
If you give, any access to your
Confluence, your Jira, et cetera,

557
00:32:13,249 --> 00:32:14,629
it'll crunch all that data.

558
00:32:14,629 --> 00:32:18,359
She will learn about all, all that
data she'll learn as well, about

559
00:32:18,359 --> 00:32:20,039
all the different chats you had.

560
00:32:20,069 --> 00:32:24,174
And if, if you tell, Annie, for
example, during a chat, oh, you

561
00:32:24,174 --> 00:32:27,254
didn't know that this service was
actually connecting to Redis, then

562
00:32:27,254 --> 00:32:32,124
she will, remember that, This service
is actually using Redis as a backend.

563
00:32:32,334 --> 00:32:37,324
So maybe next time you want to deprecate
this Redis to, to switch to Val Key.

564
00:32:37,354 --> 00:32:41,304
and you can ask any, oh, is it safe
to remove this Redis like no other

565
00:32:41,304 --> 00:32:43,814
system, I know of, are not using it?

566
00:32:43,994 --> 00:32:48,284
And she can answer you, oh, but I know
about this specific s service that

567
00:32:48,284 --> 00:32:52,094
is still using it because, I dunno,
your colleague told me that it was

568
00:32:52,094 --> 00:32:53,744
still using it like two weeks ago.

569
00:32:53,804 --> 00:32:57,334
And that's the type of things
that Annie has for the memory.

570
00:32:57,634 --> 00:33:00,094
And she also does
something like super cool.

571
00:33:00,594 --> 00:33:05,574
She spends 24 hours a day exploring the
graph, exploring your infrastructures,

572
00:33:05,574 --> 00:33:08,964
exploring your logs, exploring
your metrics, exploring new data,

573
00:33:08,964 --> 00:33:12,794
exploring your new Jira tickets,
your Linear issues, et cetera.

574
00:33:12,824 --> 00:33:19,554
And she build her own memory, her own
feelings about, maybe not feelings, but

575
00:33:20,054 --> 00:33:21,344
but feel like feeling sus.

576
00:33:21,519 --> 00:33:21,719
Yeah.

577
00:33:21,989 --> 00:33:22,139
yeah.

578
00:33:22,639 --> 00:33:27,519
But yeah, she creates some
sense of things happening how?

579
00:33:27,519 --> 00:33:29,589
Half a hour, like 24 hours a day.

580
00:33:29,589 --> 00:33:35,279
And she builds internal reports for
her and when she explore something,

581
00:33:35,279 --> 00:33:38,969
let's say, there's an incident and
she explores different hypothesis.

582
00:33:39,029 --> 00:33:43,469
And the, during one exploration,
she discovers something new,

583
00:33:43,499 --> 00:33:44,519
something she didn't know.

584
00:33:44,699 --> 00:33:46,799
She will automatically remember this.

585
00:33:46,859 --> 00:33:50,199
And if this specific thing she
discovered during an investigation

586
00:33:50,199 --> 00:33:53,999
that was useless at that moment, but
maybe tomorrow you have a, an issue

587
00:33:53,999 --> 00:33:57,304
and this is the root cause, then
in seconds she will remember it.

588
00:33:57,453 --> 00:33:59,698
I cannot do this as a human.

589
00:33:59,728 --> 00:34:03,528
Like I cannot remember all the
crap I'm saying like, every day

590
00:34:03,528 --> 00:34:05,968
by, exploring the logs and stuff.

591
00:34:05,968 --> 00:34:09,348
So it's, it's really, it's, it's
not even about tribal knowledge.

592
00:34:09,348 --> 00:34:12,858
It's really about like being
exploring 24 hours a day.

593
00:34:12,888 --> 00:34:15,108
So, I don't know, Roxane, if you
want to tell more about the memory,

594
00:34:15,158 --> 00:34:16,658
feature we have, like how it works and.

595
00:34:17,158 --> 00:34:21,098
Yes, the memory is like still a
hot topic and it's hard to handle

596
00:34:21,428 --> 00:34:24,038
because you need to understand like
what is useful and what is not.

597
00:34:24,648 --> 00:34:24,998
Right.

598
00:34:25,073 --> 00:34:28,793
You can't put every memory of
every situation all day long in,

599
00:34:28,793 --> 00:34:29,843
in the context window, right?

600
00:34:29,903 --> 00:34:32,933
Like you've gotta have some sort
of searching or, I'm guessing

601
00:34:32,933 --> 00:34:36,533
it also summarizes them so that
they're smaller and they can fit in.

602
00:34:37,033 --> 00:34:37,663
Exactly.

603
00:34:37,663 --> 00:34:39,553
So we have like different mechanism.

604
00:34:39,583 --> 00:34:43,023
So we could publish, like if you
want to read on our blog, like it's a

605
00:34:43,053 --> 00:34:46,653
state of the art with like Stanford,
we're super proud about it, how our

606
00:34:46,653 --> 00:34:51,563
agents actually learned and able to
implement the self enforcement learning

607
00:34:51,803 --> 00:34:53,813
for like the latest research paper.

608
00:34:54,313 --> 00:34:58,983
And the thing is, so what it's gonna
work is that when you see something,

609
00:34:59,073 --> 00:35:00,693
let's try to compare it to a human.

610
00:35:01,063 --> 00:35:04,753
New data on new information,
you're first going to reflect.

611
00:35:04,843 --> 00:35:08,023
So this is called the reflector
on what you've just seen,

612
00:35:08,053 --> 00:35:09,223
and do you know it already?

613
00:35:09,723 --> 00:35:12,933
And you're going to try also to
know is it something that is useful?

614
00:35:13,083 --> 00:35:15,783
So this is the first phase that
the agent will do by itself.

615
00:35:16,283 --> 00:35:19,943
And the second one, which is,
do I have some memory for that?

616
00:35:19,973 --> 00:35:21,023
And where do I put it?

617
00:35:21,523 --> 00:35:24,613
And so how do I update my
memory in a structured way?

618
00:35:25,003 --> 00:35:29,283
Because very often today, like memory
can just be markdown files, but like

619
00:35:29,283 --> 00:35:32,673
with just information all over the
place, and you don't know where to

620
00:35:32,673 --> 00:35:34,563
fetch it and what's the most important.

621
00:35:35,063 --> 00:35:38,843
So this mechanism is how do you
actually rank the information?

622
00:35:39,083 --> 00:35:44,253
It'll be done automatically by the
agents, but so how do you update and

623
00:35:44,283 --> 00:35:46,203
fetch the right memory at the right time?

624
00:35:46,703 --> 00:35:49,823
The second thing which is
tricky, is how do you handle

625
00:35:49,823 --> 00:35:51,233
short term and long term memory?

626
00:35:51,733 --> 00:35:55,963
You have short term like preferences,
things you have seen yesterday.

627
00:35:56,213 --> 00:35:57,383
I want to remember it.

628
00:35:57,443 --> 00:36:02,858
But you have also like long-term memory,
like very rare incident, very rare

629
00:36:02,858 --> 00:36:07,538
pattern that you want to have somewhere
in your brain, kind of a cold storage,

630
00:36:07,988 --> 00:36:09,458
that you want to store somewhere.

631
00:36:09,958 --> 00:36:13,918
And that if at some point it gets
really tricky, you want to go as well

632
00:36:13,918 --> 00:36:18,993
fast to the long-term memory, this
cold one, to be able to find the right

633
00:36:19,203 --> 00:36:23,363
context and then to solve the incident
or to answer to a tough question.

634
00:36:23,863 --> 00:36:26,878
So that would be the kind
of principle in place.

635
00:36:26,913 --> 00:36:28,223
I can go deeper if you want.

636
00:36:28,583 --> 00:36:28,943
Yeah.

637
00:36:29,443 --> 00:36:29,623
but

638
00:36:29,623 --> 00:36:30,013
it's,

639
00:36:30,093 --> 00:36:34,283
I was gonna say like, that sounds like
that's like the magic of a SaaS, right?

640
00:36:34,283 --> 00:36:36,413
Like that, that there's a hard problem.

641
00:36:36,913 --> 00:36:39,223
And this kinda reminds me of like
the things that people aren't talking

642
00:36:39,223 --> 00:36:42,703
about is, okay, sure, I can give it
access to infrastructure and have

643
00:36:42,703 --> 00:36:44,543
it go look at the Kubernetes API.

644
00:36:44,543 --> 00:36:46,063
Like I can have any model go do that.

645
00:36:46,063 --> 00:36:50,563
I can tell it, it has access to kubectl
and Terraform, and it can look at things

646
00:36:50,923 --> 00:36:54,793
in that moment, and it might even be able
to look at APIs and live infrastructure.

647
00:36:55,153 --> 00:37:00,383
But capturing the history of all
of our Jira tickets and all of

648
00:37:00,383 --> 00:37:03,953
our pager duty outages and the
Slack channel conversations.

649
00:37:04,253 --> 00:37:08,263
Like when I think about the big
picture of, we're really trying

650
00:37:08,263 --> 00:37:10,333
to just replace engineers, right?

651
00:37:10,333 --> 00:37:12,233
At the end of the day, we're trying
to have like a, not necessarily

652
00:37:12,233 --> 00:37:15,553
replace people, but we're trying to,
we're trying to have an additional

653
00:37:15,553 --> 00:37:21,373
AI buddy engineer that's going to
replace all the things that I forgot.

654
00:37:21,373 --> 00:37:24,523
And in this world where nobody keeps
the same job for more than a year

655
00:37:24,523 --> 00:37:28,093
or two, it's really hard to maintain
all that team knowledge, that tribal

656
00:37:28,093 --> 00:37:29,533
knowledge that isn't maybe documented.

657
00:37:29,533 --> 00:37:32,293
And none of us really enjoy
writing, handwriting, documentation.

658
00:37:32,683 --> 00:37:35,253
I think I speak, I use
one of the whisper apps.

659
00:37:35,253 --> 00:37:37,803
I don't have to even type as much
of this stuff anymore when I'm just

660
00:37:37,803 --> 00:37:39,633
sort of going off on a tangent.

661
00:37:39,633 --> 00:37:43,173
But we have this, we had this, I have
this theory that like, if these things

662
00:37:43,173 --> 00:37:48,323
are gonna eventually be someone who is
just as smart as anyone else in my team

663
00:37:48,323 --> 00:37:51,413
in terms of knowing the current state
of things, the current plans of the

664
00:37:51,413 --> 00:37:54,413
company, the current budget restrictions,
the current, you know, what do we just

665
00:37:54,413 --> 00:37:55,883
talk about in the standup this week?

666
00:37:56,163 --> 00:37:59,853
That at some point these AIs are either
gonna have to be connected in, either

667
00:37:59,853 --> 00:38:03,093
it's a two A or something, or we're
like, they're gonna need to be in the

668
00:38:03,093 --> 00:38:05,733
meetings, they're gonna need to know,
have the planning documents, they're

669
00:38:05,733 --> 00:38:09,153
gonna need to know the new budget
restrictions that we have on cloud

670
00:38:09,153 --> 00:38:10,593
compute or whatever in the company.

671
00:38:10,593 --> 00:38:12,693
Like, they're gonna need to know
all these things so that they can

672
00:38:12,693 --> 00:38:14,583
make these decisions in real time.

673
00:38:14,583 --> 00:38:18,093
Like, oh, well, you know, we have a
capacity issue, so I'm gonna spend up

674
00:38:18,093 --> 00:38:21,213
some new servers, but I also have a
budget concern, so I'm gonna use Arm,

675
00:38:21,513 --> 00:38:22,998
you know, or whatever to stay cheaper.

676
00:38:22,998 --> 00:38:26,778
Or I'm gonna, I'm gonna reduce my capacity
over here so like an increased capacity

677
00:38:26,778 --> 00:38:29,178
over here, 'cause this is optional
and I can slow that down for later.

678
00:38:29,208 --> 00:38:30,848
'cause it's just job
management or whatever.

679
00:38:30,848 --> 00:38:34,588
Like, it's gonna need a ton of
context about the business, because

680
00:38:34,588 --> 00:38:37,448
that's when I walk into teams,
that's what's happening in real

681
00:38:37,448 --> 00:38:39,158
time, is people are making decisions.

682
00:38:39,158 --> 00:38:42,488
It's not docu, it might be documented
in an email or a Slack message or

683
00:38:42,768 --> 00:38:47,378
in a document from a Zoom call, but
that's not in my infrastructure.

684
00:38:47,828 --> 00:38:50,228
That's not in my Terraform plan.

685
00:38:50,558 --> 00:38:55,658
So how does this agent, you know, maintain
the intelligence that another team

686
00:38:55,658 --> 00:38:57,338
member would maintain in the real world?

687
00:38:57,338 --> 00:38:59,778
So I'm guessing that you're
thinking that far out.

688
00:38:59,778 --> 00:39:02,598
Maybe you're thinking like a couple
years in the future when all this stuff

689
00:39:02,648 --> 00:39:05,368
is there, but you did mention already
earlier, we don't have to go back over

690
00:39:05,368 --> 00:39:08,008
it again, but like, the whole idea of
agents talking to agents, and maybe

691
00:39:08,008 --> 00:39:12,588
you have, you, maybe you have like a
meetings and email agent or something

692
00:39:12,588 --> 00:39:18,398
where you're literally CCing some sort of
agent that's digesting emails so that it,

693
00:39:18,518 --> 00:39:23,348
you know, so the epic long email threads
or slack threads are all being assessed

694
00:39:23,618 --> 00:39:27,848
in some sort of agent over here, but
then that agent over there is managing

695
00:39:27,848 --> 00:39:31,008
infrastructure and has the context of
real current status of infrastructure.

696
00:39:31,038 --> 00:39:35,428
And the two have to meet some in
some way so that your decisions

697
00:39:35,428 --> 00:39:39,898
are made in with the same awareness
that a human would have that week.

698
00:39:39,948 --> 00:39:42,888
I don't know if you're thinking of those
kind of hard problems that far out,

699
00:39:42,888 --> 00:39:45,038
but that's something where I'm trying
to imagine how we're gonna get there.

700
00:39:45,068 --> 00:39:47,348
Like, how do, how do, I'm not the smart
one that's gonna figure it out, but

701
00:39:47,348 --> 00:39:48,578
I'm just imagining that's the problem.

702
00:39:49,078 --> 00:39:52,948
I'm so excited about that because it's
really like our vision, that context

703
00:39:52,948 --> 00:39:55,918
when you speak about production, and
let's speak about only production,

704
00:39:56,158 --> 00:39:57,688
it's not only like infrastructure data.

705
00:39:58,198 --> 00:40:00,928
You also need to have like
security kind of integration,

706
00:40:01,178 --> 00:40:05,468
with for instance, identity and
ownership, business data as well.

707
00:40:05,968 --> 00:40:09,148
And when you connect all of
this data together, you cannot

708
00:40:09,148 --> 00:40:12,838
just like call different MCP
and APIs to get this context.

709
00:40:13,183 --> 00:40:17,463
I cannot insist on the, of the
fact that there's a difference

710
00:40:17,493 --> 00:40:19,563
between correlation and causality.

711
00:40:20,063 --> 00:40:23,393
When you need to make a decision,
like either, like to understand what

712
00:40:23,393 --> 00:40:26,663
is the root cause of an incident, I
like to make a change and understand

713
00:40:26,663 --> 00:40:29,603
like the best pages you need to
have the causality of events.

714
00:40:29,958 --> 00:40:32,238
How, what is the source
and what is the symptom?

715
00:40:32,738 --> 00:40:37,358
And this data, if you just take
it as MCP or APIs, it's only

716
00:40:37,358 --> 00:40:39,818
going to be like correlation
between different data sources.

717
00:40:40,318 --> 00:40:43,648
And when we speak about context
in production, you need to have

718
00:40:43,648 --> 00:40:47,308
the causality of event full
time, so that's the tough part.

719
00:40:47,728 --> 00:40:52,108
And not only for like Kubernetes,
AWS and code basis, but also for

720
00:40:52,108 --> 00:40:57,058
Slack, for Jira, for conferences,
Okta, duetta, like everything that

721
00:40:57,148 --> 00:41:01,158
makes your job like daily job kind of
interesting in terms of the context.

722
00:41:01,158 --> 00:41:04,178
You need to gather and connected
for nodes and dependencies.

723
00:41:04,678 --> 00:41:07,258
And this is the type of
context we're excited to build.

724
00:41:07,358 --> 00:41:09,368
It's not something that
you want to build yourself.

725
00:41:09,578 --> 00:41:13,408
You want to be able to make some actions
or to make decisions based on that.

726
00:41:13,798 --> 00:41:15,658
We really like focus on the context.

727
00:41:15,808 --> 00:41:17,278
How does this context look like?

728
00:41:17,778 --> 00:41:20,408
Yeah, so I guess at the end
of the day, the, Annie, right?

729
00:41:20,408 --> 00:41:22,688
Annie is the name of the AI at Anyshift.

730
00:41:23,108 --> 00:41:28,618
Like, Annie's working on the causation
while I'm sleeping, hopefully.

731
00:41:28,918 --> 00:41:34,793
Like looking at the repetitive incident
a decade ago, over a decade ago, I was

732
00:41:34,793 --> 00:41:38,223
working, with a platform that was kind of
like a Netflix platform where they were

733
00:41:38,253 --> 00:41:40,173
Argo CD in a video all around the world.

734
00:41:40,653 --> 00:41:45,353
And we had, and I was managing, this is
like pre Terraform, so we were using Salt

735
00:41:45,353 --> 00:41:51,473
Stack, I think, and we were doing, using
a lot of cloud formation and we had a bug.

736
00:41:51,813 --> 00:41:52,893
we weren't, we were the ops.

737
00:41:52,893 --> 00:41:55,733
We were the ops team, so we weren't
writing the code and the code, I think

738
00:41:55,733 --> 00:42:02,083
it was PHP code, had a a memory leak
so that we knew about once a day, this

739
00:42:02,083 --> 00:42:05,203
certain series of servers all around the
world, were gonna need to be restarted.

740
00:42:05,623 --> 00:42:10,603
And we didn't really have full like
connection failover, so it needed

741
00:42:10,603 --> 00:42:14,403
to be, you know, ideally in that
part of the world in an off hour.

742
00:42:14,403 --> 00:42:16,833
And so we had to come
up with this whole plan.

743
00:42:16,833 --> 00:42:19,893
And so for the longest time
it was just human toil.

744
00:42:20,013 --> 00:42:24,123
It was whoever was on, whoever was
available in the team at that moment was

745
00:42:24,123 --> 00:42:26,943
gonna be the one that was gonna have to
kick off a job to, for that part of the

746
00:42:26,943 --> 00:42:31,023
world and those regions to recycle all
these servers and basically reboot them.

747
00:42:31,203 --> 00:42:35,013
That was the strategy for like three
months, that was the strategy, because

748
00:42:35,013 --> 00:42:39,513
we were waiting on this supposed magic p
magic PHP fix that was gonna fix this bug.

749
00:42:39,993 --> 00:42:42,583
And this is the kind of thing where I
feel like this is the kind of thing that

750
00:42:42,583 --> 00:42:47,743
one, the AI should be just detecting, like
researching the problem and, and finding

751
00:42:47,743 --> 00:42:49,543
the fix much faster than a human would.

752
00:42:49,693 --> 00:42:55,033
And two, like I should just be able
to tell some, you know, SRE bot that,

753
00:42:55,033 --> 00:42:58,903
hey, this is the problem, automatically
create me a schedule of recycling

754
00:42:58,903 --> 00:43:02,723
these servers, until we can get the
fix that some other AI is gonna fix.

755
00:43:02,723 --> 00:43:05,513
'cause maybe the developers have
their own, you know, their own Claude

756
00:43:05,843 --> 00:43:08,483
and they're gonna, they're gonna
implement a fix eventually with that.

757
00:43:08,483 --> 00:43:12,653
But I need this solution now, and
the solution means someone's got

758
00:43:12,653 --> 00:43:15,983
to cron job this out worldwide
with all these different regions.

759
00:43:15,983 --> 00:43:19,633
And that seems like a plausible job
that I just wanna give an AI, I do not

760
00:43:19,633 --> 00:43:22,003
want to have to get up in the middle
of the night and reboot servers myself.

761
00:43:22,373 --> 00:43:25,823
And since I can't, maybe I don't
have an automation system yet,

762
00:43:25,853 --> 00:43:29,573
like, an Argo workflows or something
that can automatically schedule and

763
00:43:29,573 --> 00:43:31,043
spin things up and reboot things.

764
00:43:31,043 --> 00:43:36,943
So, I'm excited to see how this progresses
in terms of your AI being able to do,

765
00:43:37,033 --> 00:43:38,863
you know, do operations on my behalf.

766
00:43:38,983 --> 00:43:42,743
And maybe it's deterministic, I think
you mentioned earlier, you were talking

767
00:43:42,743 --> 00:43:47,473
about a script that one of your customers
needed to run, and that's also something

768
00:43:47,473 --> 00:43:50,733
recently that I've been seeing more
people experiment with as essentially

769
00:43:51,233 --> 00:43:56,413
using the AI, the non-deterministic,
crazy texting robot to write a

770
00:43:56,413 --> 00:43:59,143
deterministic workflow that it executes.

771
00:43:59,353 --> 00:44:03,703
And then eventually, you know, I think
for those of us in ops, like we, we are

772
00:44:03,703 --> 00:44:08,223
scared of the non-determinism because
we feel like that's just a crazy, no,

773
00:44:08,223 --> 00:44:12,963
not letting a crazy bot loose in my
infrastructure tore wreak havoc, havoc.

774
00:44:13,353 --> 00:44:19,323
But if I can get the AI to write a
deterministic program or workflow or

775
00:44:19,323 --> 00:44:23,363
something and then implement that,
that reduces, I think, I feel like

776
00:44:23,363 --> 00:44:27,863
the risk of the AI, because the
AI is not deciding that today I'm

777
00:44:27,863 --> 00:44:29,243
gonna do these steps out of order.

778
00:44:29,603 --> 00:44:33,713
Because that might happen, but if
it can write the code or write a

779
00:44:33,713 --> 00:44:37,103
deterministic workflow for me, that
feels like the right thing to do.

780
00:44:37,103 --> 00:44:39,263
That feels like something
a human would do.

781
00:44:39,653 --> 00:44:43,958
And I, I don't know if you're seeing
those patterns, of like people are

782
00:44:43,958 --> 00:44:48,388
using Anyshift to detect the problem
and then they're, they're not using

783
00:44:48,388 --> 00:44:51,558
AI to automate or solve the problem,
they're may be using AI to write the

784
00:44:51,558 --> 00:44:54,228
deterministic, fix or deterministic
workflow to fix that problem.

785
00:44:54,228 --> 00:44:57,223
Do you, you gave that example earlier, but
I was just curious if you had any stories

786
00:44:57,273 --> 00:45:01,233
on how that is continuing to happen or
if that's your strategy or you, how you

787
00:45:01,233 --> 00:45:03,023
recommend this to people to implement.

788
00:45:03,523 --> 00:45:03,703
Yeah.

789
00:45:03,703 --> 00:45:05,023
We actually have.

790
00:45:05,093 --> 00:45:08,893
We had a very nice story, maybe it was
two weeks ago, with that one customer.

791
00:45:09,223 --> 00:45:15,738
He so that customer got a report by Annie
saying, oh, I'm, I discovered on this

792
00:45:15,738 --> 00:45:20,168
a account, like a lot of EBS snapshots,
and it amounts to a lot of money.

793
00:45:20,588 --> 00:45:23,948
And that, that customer didn't stop here.

794
00:45:24,128 --> 00:45:26,788
He asked Annie, okay, gimme a plan.

795
00:45:26,818 --> 00:45:28,528
Like, can we do cold storage?

796
00:45:28,718 --> 00:45:31,938
Can you suggest you have access
to our Terraform stuff, so can you

797
00:45:31,938 --> 00:45:34,488
just help us just do it like now?

798
00:45:34,788 --> 00:45:40,293
And so it came, it started, by a report,
and it ended up just in in a matter

799
00:45:40,293 --> 00:45:45,133
of, minutes, to having a full fledged
solution, like a lot of money saved

800
00:45:45,433 --> 00:45:47,823
and a much better backup plan in place.

801
00:45:47,823 --> 00:45:52,443
So it's really, it's really about like
yeah, exchanging, talking with your

802
00:45:52,743 --> 00:45:57,143
agent and just having that agent the
right context to take the proper decision

803
00:45:57,143 --> 00:45:59,463
and help you in your already, hard job.

804
00:45:59,963 --> 00:46:00,563
Nice.

805
00:46:00,963 --> 00:46:04,043
I feel like we could talk about this
forever, but we, because I'm super

806
00:46:04,043 --> 00:46:07,543
interested in like the patterns in
the future of how, how we're gonna

807
00:46:07,543 --> 00:46:10,873
interface with this AI, how we're gonna
keep this AI on the rails, like how

808
00:46:10,873 --> 00:46:13,033
we're gonna keep it safe in production.

809
00:46:13,033 --> 00:46:15,013
I think there's a lot of people
that are concerned about that.

810
00:46:15,013 --> 00:46:18,003
And I think the, I'm glad to get
this episode out because I feel

811
00:46:18,003 --> 00:46:22,723
like, you know, getting the word out
on how companies are implementing

812
00:46:22,723 --> 00:46:26,953
these technologies in a reliable,
reproducible, safe way, so that we're not

813
00:46:27,453 --> 00:46:28,863
constantly tearing down infrastructure.

814
00:46:28,863 --> 00:46:32,133
You know, we all see like the
hacker news story of someone who

815
00:46:32,383 --> 00:46:35,173
had the AI run Terraform apply and
had decided to delete everything

816
00:46:35,173 --> 00:46:36,433
before it ran it again or whatever.

817
00:46:36,433 --> 00:46:37,303
That was a recent one.

818
00:46:37,363 --> 00:46:42,073
And we see those and I think sometimes
some of us were like, that's what you get.

819
00:46:42,103 --> 00:46:43,983
That's what you get for
doing that with the AI.

820
00:46:43,983 --> 00:46:45,693
Like, shame, shame on you,
you shouldn't have done that.

821
00:46:45,933 --> 00:46:48,363
But the reality is this stuff is
coming and people are experimenting

822
00:46:48,363 --> 00:46:50,583
now, like you said, you've got some
customers that are like, we want to

823
00:46:50,583 --> 00:46:53,493
go full on, like, give us the right
access, give us the implementation.

824
00:46:53,993 --> 00:46:57,113
So I think it's definitely coming and
it's exciting to talk about the different

825
00:46:57,143 --> 00:47:00,903
models, not the AI models, but the
different workflow models of how we're

826
00:47:00,903 --> 00:47:04,653
gonna do this together and safely, and how
these kind of tools are gonna integrate

827
00:47:04,653 --> 00:47:07,283
with the rest of our infrastructure
and the rest of our AI agents.

828
00:47:07,333 --> 00:47:08,593
What's, what's next?

829
00:47:08,593 --> 00:47:09,853
Like, gimme, gimme a hot take.

830
00:47:09,853 --> 00:47:12,583
What's, what's the next thing that
you're gonna, you're excited about?

831
00:47:13,083 --> 00:47:16,258
If I take this one and I think
like we have different excitement.

832
00:47:16,268 --> 00:47:17,128
Stephane and I,

833
00:47:17,383 --> 00:47:17,773
Okay.

834
00:47:18,273 --> 00:47:23,433
I would say like in term of context,
how we make it like grow even more.

835
00:47:23,933 --> 00:47:29,273
So like we want it into much more
information for the teams to like, have

836
00:47:29,333 --> 00:47:32,493
everything at hand to perform a task.

837
00:47:32,993 --> 00:47:37,863
That's, we're going to add like
quiz, Notion, like as notes in the

838
00:47:37,863 --> 00:47:40,563
graph and people are asking for it.

839
00:47:40,593 --> 00:47:43,318
I'm excited about like how big
we grow in term of context.

840
00:47:43,818 --> 00:47:44,208
Yeah.

841
00:47:44,708 --> 00:47:46,208
A giant, giant graph.

842
00:47:46,708 --> 00:47:50,348
Excited about the giant graph as well,
but I would like, like more agentic

843
00:47:50,378 --> 00:47:52,178
Annie, I want Annie to do things.

844
00:47:52,178 --> 00:47:55,178
I want Annie to proactively
like, do things.

845
00:47:55,178 --> 00:47:58,708
I will, I want my, I, I want
AGI, I want to sleep at night.

846
00:47:58,708 --> 00:48:03,678
So I want the, I want Annie to detect
problems, fix problems, and just

847
00:48:03,678 --> 00:48:05,188
send me an email tomorrow morning.

848
00:48:05,458 --> 00:48:07,138
That's, that's the shot at roadmap.

849
00:48:07,168 --> 00:48:08,608
I'm really excited about this.

850
00:48:09,108 --> 00:48:09,528
Yeah.

851
00:48:09,528 --> 00:48:13,188
I do like the idea of like getting up
in the morning, I'm brushing my teeth

852
00:48:13,218 --> 00:48:17,718
and I've got my Annie app on my phone,
and we're having a voice chat, kinda

853
00:48:17,718 --> 00:48:21,143
like, I just did the, this morning
with ChatGPT, where we're having a

854
00:48:21,143 --> 00:48:23,828
conversation and I'm like, okay, what
happened overnight while I was sleeping?

855
00:48:23,888 --> 00:48:26,978
Like, gimme, gimme, like, did we
have unusual spikes in traffic?

856
00:48:26,978 --> 00:48:28,508
Did we have any sort of hiccups?

857
00:48:28,508 --> 00:48:30,728
Was there any cloud outages
while I was sleeping?

858
00:48:31,058 --> 00:48:33,218
You know, like we all wake up
and GitHub's not running today.

859
00:48:33,218 --> 00:48:35,488
You know, that's the kind of
world that we live in right now.

860
00:48:35,488 --> 00:48:40,548
And instead of me having to like peruse
Hacker news or check status pages

861
00:48:40,548 --> 00:48:43,238
of things or, you know, everything
might be fine right now, but there

862
00:48:43,238 --> 00:48:45,548
might have been a ton of stuff
that happened in the last 10 hours.

863
00:48:45,938 --> 00:48:50,558
And instead of me having to read Slack,
like just infinite slack messages and

864
00:48:50,558 --> 00:48:53,978
channels that just are giving me alert
fatigue all day long, I'd love to

865
00:48:53,978 --> 00:48:57,683
really have, like morning summaries,
that are, that essentially read to

866
00:48:57,683 --> 00:49:03,333
me by either a British or Australian
accent AI that because I gotta have a

867
00:49:03,513 --> 00:49:04,453
the killer feature.

868
00:49:04,513 --> 00:49:06,023
Let's let's strip this.

869
00:49:06,253 --> 00:49:07,693
I got a French accent over here.

870
00:49:07,693 --> 00:49:10,033
I got a South African accent over here.

871
00:49:10,033 --> 00:49:13,398
Like, I just, I want all the, I want
all the robots to have more personality

872
00:49:13,398 --> 00:49:16,748
so that I, I recognize their voice when
we're in a group call with nothing but me

873
00:49:16,748 --> 00:49:20,288
and the agents, and one of them knows my
infrastructure and one of them knows my

874
00:49:20,288 --> 00:49:22,518
GitHub situation and, you know, whatever.

875
00:49:22,518 --> 00:49:24,508
One of them's cooking
my breakfast downstairs.

876
00:49:24,508 --> 00:49:25,438
I don't know what's gonna happen.

877
00:49:25,438 --> 00:49:28,338
But it's exciting to see this
future unfolding in front of us.

878
00:49:28,338 --> 00:49:32,668
And I think, the last thing here is
like, when is that a year from now?

879
00:49:32,668 --> 00:49:34,018
is that five years from now?

880
00:49:34,018 --> 00:49:38,158
Like where do you, can you even imagine
where Anyshift is gonna be in a year?

881
00:49:38,158 --> 00:49:41,338
Do you even have a vision board of
where this is all going in a year?

882
00:49:41,838 --> 00:49:45,528
I don't know how fast the transition
from agent to agent will happen.

883
00:49:45,828 --> 00:49:49,608
For sure, like today, like
incident management and resolution

884
00:49:49,938 --> 00:49:51,318
is still an unsolved problem.

885
00:49:51,318 --> 00:49:53,958
For instance, it'll not
be manual in the future.

886
00:49:54,458 --> 00:49:58,148
Six now, six months, a year
from now, it should be solved.

887
00:49:58,448 --> 00:50:03,698
And like how agents actually also
like perform action in production

888
00:50:04,198 --> 00:50:08,538
should be something that I believe
in the six months, a year from

889
00:50:08,538 --> 00:50:13,318
that, will, will be done as like
the models get so much better.

890
00:50:13,818 --> 00:50:17,808
We made a bet from day one that
we would be the best type context

891
00:50:18,018 --> 00:50:19,488
and that agents will improve.

892
00:50:19,818 --> 00:50:25,218
And we have seen the leap of like,
progress, like Sonet 4.6, for instance.

893
00:50:25,698 --> 00:50:29,328
It's only gonna get better and
we are the best at providing this

894
00:50:29,328 --> 00:50:30,858
causality chain between events.

895
00:50:31,128 --> 00:50:34,398
And then our agent will be able
to be the best to solve it with

896
00:50:34,398 --> 00:50:36,168
the latest models improving.

897
00:50:36,668 --> 00:50:36,938
Yeah.

898
00:50:36,938 --> 00:50:37,988
A new model to me.

899
00:50:37,988 --> 00:50:40,238
I mean, I feel like we're all
in the lull right now waiting

900
00:50:40,238 --> 00:50:43,488
for the, waiting for whoever's
gonna release their next version.

901
00:50:43,968 --> 00:50:48,483
And it feels like right now, every time
we have a major Sota model release a

902
00:50:48,483 --> 00:50:52,983
new version, it makes things that we
thought we needed to do in our tooling,

903
00:50:53,483 --> 00:50:57,523
irrelevant, because the model can now
do that now, or, you know, the model's

904
00:50:57,523 --> 00:50:59,353
been updated so that I don't need that.

905
00:50:59,563 --> 00:51:02,173
I don't need for it to read the
documentation every time because now

906
00:51:02,173 --> 00:51:03,763
it's been trained on that documentation.

907
00:51:03,763 --> 00:51:06,263
before it was, a lot of the models
I think we're using now, were built,

908
00:51:06,503 --> 00:51:08,093
were built in September last year.

909
00:51:08,093 --> 00:51:10,963
So, a lot of the tools I'm
using barely even exist.

910
00:51:10,963 --> 00:51:12,763
I mean, OpenClaw wasn't
even a thing back then.

911
00:51:12,763 --> 00:51:14,223
Like, there's so many things.

912
00:51:14,253 --> 00:51:15,693
Docker sandboxes didn't exist.

913
00:51:15,693 --> 00:51:18,063
Like, we have so many new things
that, that models don't know about.

914
00:51:18,063 --> 00:51:20,973
So I have to constantly point it
to websites, give it documentation.

915
00:51:21,473 --> 00:51:23,183
So I'm excited for that future too.

916
00:51:23,183 --> 00:51:26,093
And and it looks like, even though
I can't predict what's gonna happen

917
00:51:26,093 --> 00:51:27,533
in six months, I feel like the,

918
00:51:28,033 --> 00:51:30,028
it's all gonna be amazing
and I'm here for it.

919
00:51:30,028 --> 00:51:31,318
So that's why we had these podcasts.

920
00:51:31,773 --> 00:51:34,223
it's great to have you on
the, podcast and I'm looking

921
00:51:34,223 --> 00:51:35,693
forward to what you all do next.

922
00:51:36,193 --> 00:51:37,058
Thank you so much.

923
00:51:37,558 --> 00:51:38,053
Thank you so much,

924
00:51:38,068 --> 00:51:38,458
Alright.

925
00:51:38,668 --> 00:51:39,258
You can find Anyshift@anyshift.io.

926
00:51:41,238 --> 00:51:44,108
I'm assuming both of you, should
people follow you on LinkedIn?

927
00:51:44,108 --> 00:51:46,728
I guess where are the socials
that people should find you all?

928
00:51:47,228 --> 00:51:47,798
Twitter and LinkedIn.

929
00:51:48,218 --> 00:51:48,508
yeah.

930
00:51:48,578 --> 00:51:49,238
Twitter, LinkedIn.

931
00:51:49,738 --> 00:51:50,158
Nice.

932
00:51:50,248 --> 00:51:50,638
All right.

933
00:51:51,058 --> 00:51:51,688
Ciao everybody.

934
00:51:52,457 --> 00:51:54,797
Thanks for joining us, and I'll
see you in the next episode.