WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=Byv311hdoHE

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/Byv311hdoHE.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.560 --> 00:00:13.520
Thanks guys for coming.

s3
00:00:13.520 --> 00:00:18.240
I know right after lunch, this is like prime time like sleepiness.

s4
00:00:18.240 --> 00:00:19.960
I'm curious just to get a sense for the audience.

s5
00:00:19.960 --> 00:00:22.320
Who here is a forward deployed engineer?

s6
00:00:22.320 --> 00:00:25.440
Okay, who here is hiring for a deployed engineers?

s7
00:00:25.440 --> 00:00:26.440
Okay, helpful.

s8
00:00:26.440 --> 00:00:27.400
Anyone else in between?

s9
00:00:27.400 --> 00:00:29.560
You can raise your hand just like if if I didn't cover it.

s10
00:00:29.560 --> 00:00:30.400
Okay.

s11
00:00:30.400 --> 00:00:31.040
I see you.

s12
00:00:31.040 --> 00:00:32.680
Double double hand raise.

s13
00:00:32.680 --> 00:00:34.160
Um well, I'm Natalie Mier.

s14
00:00:34.160 --> 00:00:36.920
I am the um head of agent engineering at Sierra.

s15
00:00:36.920 --> 00:00:38.400
I've been at the company about 2 years.

s16
00:00:38.400 --> 00:00:40.640
I'll tell you a little bit about that.

s17
00:00:40.640 --> 00:00:42.560
But honestly, I'm here to talk about forward deployed engineering.

s18
00:00:42.560 --> 00:00:53.320
So, we're actually going to spend a little bit less time talking about Sierra and a lot more time talking about the history of forward deployed engineering and and the dirty secret, I like to call it, of the domain.

s19
00:00:53.320 --> 00:00:55.920
And you all can tell me how controversial it is.

s20
00:00:55.920 --> 00:01:03.440
So, um just a little bit about me, which will give you a sense for sort of how I'm coming at this domain, at the whole domain of forward deployed engineering.

s21
00:01:03.440 --> 00:01:05.840
I actually started off as a policy nerd.

s22
00:01:05.840 --> 00:01:07.200
So, I was at Georgetown.

s23
00:01:07.200 --> 00:01:09.480
Um I studied uh in the School of Foreign Service there.

s24
00:01:09.480 --> 00:01:12.440
And so, I was obsessed with basically tech policy.

s25
00:01:12.440 --> 00:01:22.360
I um learned to code and then swindled my way or earned my way, depending upon your perspective, into a job at Palantir, where I spent 2016 to 2021.

s26
00:01:22.360 --> 00:01:30.160
And um there I worked on the privacy team and I was a an infrastructure engineer and also a forward deployed engineer with law enforcement and defense primarily.

s27
00:01:30.160 --> 00:01:39.280
So, I I've seen forward deployed engineering up close and I'm going to talk a lot about the history of not just my time at Palantir, but actually the time that predated me to talk about basically the origins

s28
00:01:39.280 --> 00:01:41.840
of the entire discipline.

s29
00:01:41.840 --> 00:01:45.800
And then I actually wanted to get a business degree of all things, so I went to Stanford.

s30
00:01:45.800 --> 00:01:49.560
And now I'm at Sierra from 24 to to 26.

s31
00:01:49.560 --> 00:01:57.960
And one of the things that I think is is quite interesting and when I started at Sierra, actually a little known fact is that I joined the deployment team.

s32
00:01:57.960 --> 00:02:00.320
and at the time I hated the name.

s33
00:02:00.320 --> 00:02:12.120
I I thought this is not what we're doing and B it sort of is a nonsensical name so to speak and so I actually wrote an an article in July of 2024, so about 6 months after I started,

s34
00:02:12.120 --> 00:02:16.280
um about 2 years ago now, on this concept of the agent engineer.

s35
00:02:16.280 --> 00:02:22.920
And the idea was that agent engineering was actually a subdiscipline of AI engineering which we're of course here to talk about.

s36
00:02:22.920 --> 00:02:31.600
And the vision was that agent engineering had the same level of customer accountability as forward deployed engineering but was actually a domain onto itself.

s37
00:02:31.600 --> 00:02:40.520
And I think since then, you know, in the past 2 years we've actually seen a lot of folks talking about agent engineering as a discipline and now we're actually seeing harness engineering as sort of a new discipline

s38
00:02:40.520 --> 00:02:42.520
sort of is almost a subset of agent engineering.

s39
00:02:42.520 --> 00:02:45.880
And so if you're confused, that's sort of what this talk is about.

s40
00:02:45.880 --> 00:02:53.160
And so the dirty secret of forward deployed engineering, I won't make you wait until the end, is it doesn't exist.

s41
00:02:53.160 --> 00:03:00.200
It is a term that it is meant to describe so many things that it sort of means nothing at this point.

s42
00:03:00.200 --> 00:03:02.720
But nonetheless, it's sort of the hottest job in AI.

s43
00:03:02.720 --> 00:03:12.800
And so what I hope to convince you of over the next, let's say, 15 minutes is A that this is true and B that this doesn't matter.

s44
00:03:12.800 --> 00:03:19.040
And so um I'm going to dive in and we're going to start with tracing the history of forward deployed engineering from Palantir.

s45
00:03:19.040 --> 00:03:20.840
So we're going to start in 2008.

s46
00:03:20.840 --> 00:03:32.440
We're going to basically come back to the present and then I'll share with you at least my perspective on what forward deployed engineering is today, why you should want to or not want to be a forward deployed engineer, and why we're actually all forward deployed engineers

s47
00:03:32.440 --> 00:03:33.600
in the end.

s48
00:03:33.600 --> 00:03:39.080
Um so first off, I I'd actually like to start in 2016 and then we'll go a little back in time.

s49
00:03:39.080 --> 00:03:42.519
And I actually pulled basically Palantir job postings in 2016.

s50
00:03:42.519 --> 00:03:53.440
This is when I started and what drew me into Palantir was effectively this concept of a forward deployed software engineer and this was an emphasis on the forward deployed part.

s51
00:03:53.440 --> 00:03:54.280
I mean, they're leading.

s52
00:03:54.280 --> 00:03:58.720
These days you lead with, you know, you speak a certain language, you have these skill sets.

s53
00:03:58.720 --> 00:04:00.280
They were leading with location.

s54
00:04:00.280 --> 00:04:06.680
And there's a reason for that that we'll get to, but but forward-deployed literally meant sitting with customers, being on the ground.

s55
00:04:06.680 --> 00:04:14.440
And in part, as we'll talk about, this was because so much of Palantir's um ecosystem was on-prem deployments.

s56
00:04:14.440 --> 00:04:18.720
And so, when we think of forward-deployed engineering though, we kind of think of this.

s57
00:04:18.720 --> 00:04:29.080
And so, um this is sort of the prototypical forward-deployed engineer right out of college um parachuting out of a helicopter um and of course, you know, fixing some sort of runtime issue

s58
00:04:29.080 --> 00:04:30.760
um in the air.

s59
00:04:30.760 --> 00:04:36.919
But in practice, forward-deployed engineering actually started as something more akin to DevOps.

s60
00:04:36.919 --> 00:04:41.960
And so, platform stability really constituted the bulk of the job early on.

s61
00:04:41.960 --> 00:04:51.360
And so, when you think about um what forward-deployed engineering actually was, don't worry about understanding this code on the right, but it just gives you a sense for what what folks were actually doing.

s62
00:04:51.360 --> 00:04:58.360
In fact, my onboarding project at Palantir was actually deploy the software on an EC2 instance.

s63
00:04:58.360 --> 00:05:00.520
And so, that was literally the job.

s64
00:05:00.520 --> 00:05:11.480
And it was a a large part of the job, and so much so that um basically the job felt like this, which is basically an email from the customer telling you that someone mistakenly unplugged

s65
00:05:11.480 --> 00:05:12.720
um the instance again.

s66
00:05:12.720 --> 00:05:13.680
Hey, could you look into it?

s67
00:05:13.680 --> 00:05:16.840
By the way, it's 2:00 a.m. and you need to go in.

s68
00:05:16.840 --> 00:05:25.360
Um and so, in practice, a lot of the early forward-deployed engineering was was really focused on DevOps.

s69
00:05:25.360 --> 00:05:30.360
And come 2012, the the Palantir platform got a lot more stable.

s70
00:05:30.360 --> 00:05:32.919
And Palantir built data integration software.

s71
00:05:32.919 --> 00:05:38.120
So, quick show of hands, everyone here probably knows, but who here knows what data integration software is?

s72
00:05:38.120 --> 00:05:38.480
Great.

s73
00:05:38.480 --> 00:05:42.720
Anyone think data integration software is really useful without data?

s74
00:05:42.720 --> 00:05:43.280
Okay.

s75
00:05:43.280 --> 00:05:43.760
Oh, okay.

s76
00:05:43.760 --> 00:05:44.760
One All right.

s77
00:05:44.760 --> 00:05:46.760
I'm curious to hear your take after.

s78
00:05:46.760 --> 00:05:54.360
Um So, um so I like to think of data integration software without integrated data as a movie theater that's playing nothing.

s79
00:05:54.360 --> 00:05:56.840
It's like, why show up, you know?

s80
00:05:56.840 --> 00:06:00.360
And so, this is basically what Palantir was selling without forward deployed engineers.

s81
00:06:00.360 --> 00:06:03.080
A, the movie theater wasn't standing in 2008.

s82
00:06:03.080 --> 00:06:06.480
By 2012, you actually don't have data in it, right?

s83
00:06:06.480 --> 00:06:08.280
You don't have movies playing.

s84
00:06:08.280 --> 00:06:15.440
And so, you sort of have this basic, um you know, comic explaining this this process, which is, "Hey, great.

s85
00:06:15.440 --> 00:06:17.720
We have this data integration software, isn't it useful?"

s86
00:06:17.720 --> 00:06:24.760
Right, but my data's in 20 places, and now you actually have a forward deployed engineer that needs to go in and integrate that data.

s87
00:06:24.760 --> 00:06:28.520
And most of this was written in, you know, variants of Java back in the day.

s88
00:06:28.520 --> 00:06:34.320
Um as was actually Palantir's client, fun fact, in Java Swing, if anyone here has had the privilege to to use that.

s89
00:06:34.320 --> 00:06:39.440
Um So, so basically, forward deployed engineers now have this dual role, right?

s90
00:06:39.440 --> 00:06:42.480
DevOps plus data integration.

s91
00:06:42.480 --> 00:06:48.600
And so, they wanted to to deeply understand the customer's environment to to model that data appropriately.

s92
00:06:48.600 --> 00:06:56.000
And what Palantir uh did then and now still does call an ontology, which is sort of a famous Palantir um term of art.

s93
00:06:56.000 --> 00:06:59.320
But you can think of it as a taxonomy for that data.

s94
00:06:59.320 --> 00:07:02.760
And then 2016 rolled around, and you have custom solutions.

s95
00:07:02.760 --> 00:07:07.640
So, folks are building effectively solutions to known problems.

s96
00:07:07.640 --> 00:07:11.160
Um and usually that took the form of a Slate dashboard.

s97
00:07:11.160 --> 00:07:13.000
Has anyone here heard of Slate?

s98
00:07:13.000 --> 00:07:14.880
This was sort of an internal Okay.

s99
00:07:14.880 --> 00:07:15.480
All right.

s100
00:07:15.480 --> 00:07:18.880
Um so, Slate was basically um a drag-and-drop builder.

s101
00:07:18.880 --> 00:07:21.440
It actually still exists as part of Palantir's platform.

s102
00:07:21.440 --> 00:07:27.000
It allowed you to map, similar to Retool actually, um map components on the page to data sources.

s103
00:07:27.000 --> 00:07:31.240
So, now you have the data, and then the question became, "Okay, how do we make it useful?"

s104
00:07:31.240 --> 00:07:41.560
And the the person or the individual or the group that was most well situated to make that data useful was the the individuals that most understood that data, which were the forward deployed engineers.

s105
00:07:41.560 --> 00:07:50.760
And so, um the this basically era of of Slate actually gave way to an era of Foundry, which is Palantir's main platform today.

s106
00:07:50.760 --> 00:07:55.200
And the era of Foundry was effectively all about um making data useful.

s107
00:07:55.200 --> 00:08:01.000
So, Palantir had this phrase of uh going from data to decision-making, which was was sort of the core problem.

s108
00:08:01.000 --> 00:08:09.360
One of the things they found, a spoiler alert for anyone building a dashboarding platform, is that a dashboard that doesn't write back to the data source isn't that useful.

s109
00:08:09.360 --> 00:08:13.720
That these things would decay over time and actually not be as useful as they could be.

s110
00:08:13.720 --> 00:08:18.480
Fast forward to 2020, you know, Palantir is on the verge of IPO.

s111
00:08:18.480 --> 00:08:28.200
And they also need a solution that doesn't involve shipping people out to Finland or Canberra or choose the kind of location on the earlier slide of your preference.

s112
00:08:28.200 --> 00:08:33.200
And um they actually wanted to empower customers to do more of this work themselves.

s113
00:08:33.200 --> 00:08:35.960
And so, this is where Palantir really started to become a platform.

s114
00:08:35.960 --> 00:08:41.360
And the job of forward deployed engineering still involved the first three, but it also involved enablement of those customers.

s115
00:08:41.360 --> 00:08:51.360
So, if you look at the work that Palantir's done with Airbus, for example, in Skywise, that was focused on enabling thousands of Airbus engineers on the Foundry platform to do work

s116
00:08:51.360 --> 00:08:52.680
of the forward deployed engineer.

s117
00:08:52.680 --> 00:09:01.080
And these days, they actually have a platform uh since I left called AIFTE that's focused on leveraging uh large language models to do similar work.

s118
00:09:01.080 --> 00:09:04.120
So, this is the timeline, and you'll see that it's not one thing.

s119
00:09:04.120 --> 00:09:14.480
And so, this is a a picture of uh Alex Karp, and I think we might have lost the photo um over over here, but of Alex Karp basically running an AIP

s120
00:09:14.480 --> 00:09:15.640
bootcamp.

s121
00:09:15.640 --> 00:09:22.600
So, this is basically where the forward deployed engineers or the deployment strategists would go in and actually teach customers to use the software.

s122
00:09:22.600 --> 00:09:28.280
So, what's interesting about this is these weren't discrete eras, right?

s123
00:09:28.280 --> 00:09:31.800
The total FTE jobs to be done actually increased over time.

s124
00:09:31.800 --> 00:09:35.720
Maybe they peaked at one point, but they actually grew over time.

s125
00:09:35.720 --> 00:09:46.120
And so, now we have a world come 2020, 2024 where the role of FDE sort of doesn't mean anything quite yet or it means everything, right?

s126
00:09:46.120 --> 00:09:48.880
It's actually the best training ground for generalists.

s127
00:09:48.880 --> 00:09:59.600
And in fact, this is why I would say that Palantir has seen such success Palantir alums founding companies is because it's actually training you on all of these different facets of the role.

s128
00:09:59.600 --> 00:10:07.640
And so I have an idea that I like to think of that you all can ask your next candidate if you're interviewing for for engineering role.

s129
00:10:07.640 --> 00:10:10.480
Which is what vintage of FDE are you?

s130
00:10:10.480 --> 00:10:16.320
Are you the 2008 vintage of platform stability with hints of panic.

s131
00:10:16.320 --> 00:10:21.400
2012, 2016 2020 or even you know today.

s132
00:10:21.400 --> 00:10:24.440
What is what is the forward deployed engineer of today?

s133
00:10:24.440 --> 00:10:35.280
And so that I like to call this the FDE vintage and you might debate the years and debate the details, but just like fine wine, I think there's a particular vintage of an FDE.

s134
00:10:35.280 --> 00:10:36.880
So welcome to 2026.

s135
00:10:36.880 --> 00:10:40.600
Here we are and FDE is absolutely everywhere.

s136
00:10:40.600 --> 00:10:48.880
So I'm sure you all have heard Google recently announced with GCP an effort to hire customer engineers or FDEs.

s137
00:10:48.880 --> 00:10:54.960
Open AI created a new unit with a massive investment to aid the corporate AI push.

s138
00:10:54.960 --> 00:11:02.800
You have folks like Aaron Levie basically posting on X as as well as many others about just how important this work is today.

s139
00:11:02.800 --> 00:11:07.880
But if you think as about what we just saw, the work isn't one thing, right?

s140
00:11:07.880 --> 00:11:11.280
So forward deployed engineering is actually many things all at once.

s141
00:11:11.280 --> 00:11:19.560
And in fact, today if you look at basically what it takes to be a forward deployed engineer, you were to combine all the job postings into one, you'd probably see something like this.

s142
00:11:19.560 --> 00:11:27.600
You know, I don't know how many of you have been a staff eng for eight years with six years of direct sales experience and four years as a solution architect.

s143
00:11:27.600 --> 00:11:31.480
Also, I don't know if you've taught in schools prior because that would help as well.

s144
00:11:31.480 --> 00:11:38.960
And this is actually what it feels like hiring for these deployed roles as I've done for the past, you know, 2 years, 2 and 1/2 years or so.

s145
00:11:38.960 --> 00:11:44.480
Um, but this is actually what's being asked of forward deployed engineers today.

s146
00:11:44.480 --> 00:11:52.760
And so you think about the the dirty secret that it doesn't exist and actually it it either doesn't exist or maybe it actually exists too much.

s147
00:11:52.760 --> 00:11:53.920
Hard to say.

s148
00:11:53.920 --> 00:12:07.880
Um, and part of this, I think, and part of sort of um, I think the importance of the role is that across all of those different aspects of the role, the one continuity point is that you actually have

s149
00:12:07.880 --> 00:12:16.360
every single forward deployed engineer accountable to the customer, whether that be for devops, for enablement, whether that be for custom solutioning, data integration.

s150
00:12:16.360 --> 00:12:21.440
These are all forms of solutions or customer accountability.

s151
00:12:21.680 --> 00:12:31.080
And when that happens though, I I ask the question, you know, what happens to engineering broadly, even outside of forward deployed engineering, when code becomes cheap to produce?

s152
00:12:31.080 --> 00:12:38.160
When it becomes really quick to fire off a prompt to a background agent and get something pretty great on the other end.

s153
00:12:38.160 --> 00:12:47.680
And one way to think about this is forward deployed engineers being accountable to customers and trying to translate that signal into the product, either to make it more stable

s154
00:12:47.680 --> 00:12:50.040
or to drive more data integration.

s155
00:12:50.040 --> 00:12:58.440
And so what this allows you to do as a forward deployed engineer today is actually be way more impactful in product development.

s156
00:12:58.440 --> 00:13:06.800
So forward deployed engineers can now not just talk to customers, not just prototype, but actually build end-to-end solutions with coding agents.

s157
00:13:06.800 --> 00:13:12.040
But I think something more interesting is happening that we're seeing at Sierra, which is actually the lines are blurring.

s158
00:13:12.040 --> 00:13:15.960
So product engineering is also becoming more client-facing.

s159
00:13:15.960 --> 00:13:24.040
And so if you're a good product engineer, if you're a good forward deployed engineer, you should be thinking about the product and the customer both together.

s160
00:13:24.040 --> 00:13:33.080
And so one of the reasons that folks are talking about forward deployed engineering as so essential to where we are at this point in time is that it's actually converging.

s161
00:13:33.080 --> 00:13:37.480
Forward-deployed engineering is in some ways actually getting larger than we ever thought it was before.

s162
00:13:37.480 --> 00:13:39.800
So, it's actually stacking more skills on top.

s163
00:13:39.800 --> 00:13:44.600
And if you're a product engineer, even an infra engineer, you should also be thinking about DevOps, right?

s164
00:13:44.600 --> 00:13:47.600
How do you actually deploy the software?

s165
00:13:48.440 --> 00:13:52.200
And then separately, a different shift is changing when code becomes cheap.

s166
00:13:52.200 --> 00:14:00.760
And this is actually from um both Emergence Capital and then a a blog post from someone on on our team, our head of go-to-market ops, Elliot Greenwald.

s167
00:14:00.760 --> 00:14:03.840
And the way that we sell software is changing.

s168
00:14:03.840 --> 00:14:08.720
And at Sierra, we've always been focused on this outcome-based pricing model.

s169
00:14:08.720 --> 00:14:14.040
We think that you should pay a software platform for the value that that software delivers.

s170
00:14:14.040 --> 00:14:15.440
And this is not always possible.

s171
00:14:15.440 --> 00:14:23.920
So, if you think about uh scenarios where basically you can only slightly attribute the outcome to the product, think of seat-based pricing.

s172
00:14:23.920 --> 00:14:29.080
This is basically the way that software has been priced for uh an incredibly long time.

s173
00:14:29.080 --> 00:14:32.440
And then you think about the agency and the autonomy to achieve the outcome.

s174
00:14:32.440 --> 00:14:35.080
And agents basically move us up into the right here.

s175
00:14:35.080 --> 00:14:42.839
So, usage-based, a good example of that might actually be what we pay to OpenAI or Anthropic or some of the foundation model providers.

s176
00:14:42.839 --> 00:14:46.600
Because it's it's based on usage and it's hard to actually attribute the outcome.

s177
00:14:46.600 --> 00:14:50.520
When you think about customer experience in AI, that's outcome-based.

s178
00:14:50.520 --> 00:14:52.440
It's are you actually making a sale?

s179
00:14:52.440 --> 00:14:54.760
Are you solving a customer's inquiry?

s180
00:14:54.760 --> 00:15:00.640
And we think basically that most of pricing in this market will move to outcome-based.

s181
00:15:00.680 --> 00:15:09.480
So, if you have both of these things, you have forward-deployed engineers that can now contribute to the product, you have more outcome-based pricing, how do you actually guarantee the outcome?

s182
00:15:09.480 --> 00:15:12.120
And that really is forward-deployed engineering.

s183
00:15:12.120 --> 00:15:20.360
And so, if I jump here uh back to sort of where I started, this is how I envision agent engineering in in 2024.

s184
00:15:20.360 --> 00:15:26.240
And I I don't know if I was wrong or you guys can tell me if I was wrong, but it wasn't the whole story.

s185
00:15:26.240 --> 00:15:26.840
Right?

s186
00:15:26.840 --> 00:15:30.760
That agent engineering actually was a subset heavily focused on outcomes.

s187
00:15:30.760 --> 00:15:33.360
It is a flavor of forward deployed engineering.

s188
00:15:33.360 --> 00:15:43.360
But these days, I'm actually of a different mind, which is that um really in some ways everything is forward deployed engineering, or at least everything is trending that way.

s189
00:15:43.360 --> 00:15:49.520
Product engineering, agent engineering, AI engineering, solutions engineering, and customer engineering.

s190
00:15:49.520 --> 00:15:56.400
When you think about enablement, you think about devops, you think about solution building, data integration, building an agent, deploying it into production.

s191
00:15:56.400 --> 00:15:59.600
These are things that are on behalf of customers at the end of the day.

s192
00:15:59.600 --> 00:16:02.360
And we should be pricing outcomes associated with them.

s193
00:16:02.360 --> 00:16:11.480
And forward deployed engineering as a concept, even if it is one, sort of lacks a coherent definition, is something that actually allows us to enable those outcomes.

s194
00:16:11.480 --> 00:16:17.320
So, I'll leave you with um maybe one last note, which is that forward deployed engineering is dead.

s195
00:16:17.320 --> 00:16:19.480
And long live forward deployed engineering.

s196
00:16:19.480 --> 00:16:20.400
So, thank you so much.

s197
00:16:20.400 --> 00:16:24.800
And by the way, we're hiring across forward deployed engineering roles uh at Sierra.

s198
00:16:24.800 --> 00:16:29.000
So, I told you it wouldn't be about Sierra, but if you want to talk about FDE, I'm here.

s199
00:16:29.000 --> 00:16:31.026
Thanks, guys.

s200
00:16:31.026 --> 00:16:33.026
[applause]

s201
00:16:46.746 --> 00:16:48.746
[music]