# Scaling Agentic AI: Airbnb's 97% Adoption Blueprint

**Podcast:** Engineering Enablement by DX
**Published:** 2026-06-22

## Transcript

Welcome back to the Engineering Enablement Podcast.
I'm your host, Justin Riat.
This episode was recorded live at DX Annual, where we were joined by Christopher Sanson and Madison Capps from Airbnb, where they're leading the development of the company's internal AI platform and pushing agentic AI well beyond the IDE.
In this session, they share how Airbnb reached 97% weekly adoption of agentic AI among engineers without a single mandate and saw PR throughput jump 65% as a result.
They break down the internal product, AirChat, that made it all possible, how non-developers across finance, design, and operations organically started using coding agents in their daily work, and why Airbnb is now building towards asynchronous AI, where developers manage multiple agents running in the background.
Thanks for listening.
Nice to meet everybody.
My name's Christopher.
This is Madison.
I'm going to kick us off, talk a little bit about what we've been doing with AI at Airbnb, and then Madison.
is going to come on and actually talk about what's actually been happening from an engineering perspective.
All right.
So we're going to talk a little bit about what we're doing with non-developers and sort of the next wave of what's next, along with like some asynchronous AI stuff.
But I don't know, like the vibe, it's kind of been doom and gloom so far.
So far, like, I don't know, this morning was like, you know, you read the news and it's like, AI is taking our jobs.
And then AI actually isn't making us that more productive.
And like, by the way, those are complete opposite sentiments.
So I don't know which one is true, but like.
I don't know, it costs too much.
And like, we need mandates for people to adopt it and things like that.
And I read these headlines and stuff like that.
And I think what we're seeing inside of Airbnb is kind of the opposite.
Like we're seeing it unlock like new levels of productivity.
These are like the most wildly successful tools we've ever rolled out before.
So I wanted to take a minute about, to talk about a little bit about what we're seeing in, I'd say Airbnb and like actually share some numbers with you.
And I want to do it by going through like a couple of myths that you do see in the news a lot that I think, you know, have kind of been true, but I think this is all going to change very, very fast because like, especially since like December, January, probably like a lot of you, a lot of the old metrics have just completely changed.
So I kind of want to walk through a little bit of that with you all today.
And the first thing that I want to really tackle is that like AI, the value of AI is all about replacing humans.
Actually, I loved what Jennifer was talking about this morning around like setting the tone and the culture around AI tools, like feeds into everything else that's happening in terms of adoption, in terms of utilization.
And again, I don't know about you guys, I've been on developer platform for a long time, and our CFO was never asking us what we were doing.
You know what I mean?
Finance did not care about PR throughput, but now all of a sudden they're sort of, hey, what's going on with AI productivity?
We're like, why are you asking us this?
But people can kind of check that vibe.
I think the way I like to think about it is like, You know, Steve Job has this great quote about the computer, right?
Some of you probably know, like he called it a bicycle for the mind, you know?
A lot of people probably know that quote.
There's actually, I don't know if you know the full story though, behind it.
There was a little more to it.
He actually talked about it in the context of a study.
There was a scientific study that looked at the efficiency of motion, like locomotion.
Like how much energy did animals take to move one kilometer versus humans?
And what they found was like animals were way more efficient.
Humans were like dead last.
Like the condor uses way less energy to travel one kilometer than a person.
I don't know why they had to do a study.
They could like I could have just told them that.
Like they don't I don't know why they need to figure that out.
But they did it.
And then what happened is that somebody said, OK, you know, that makes sense.
But what if we put a human on a bicycle and did that same comparison?
And a human on a bike like blew the animals away.
It was like first it went from last to first.
And that was his context.
He said computers are like a bicycle for the mind.
And I love it.
And I think we're seeing that all over again.
Like we like to call it sort of like, if that's a bicycle, like AI is kind of like a motorcycle for the mind.
We're seeing people achieve levels of sort of creativity, unlock and productivity that really wasn't possible before.
The second mandate is, the second myth is around, I think, mandates to drive AI adoption, right?
So, you know, this is, you know, you probably all read like Promatic Engineer, you read the DX blogs.
you know who they are about like measuring token usage and performance reviews.
And I think it was really encouraging that most of the folks this morning have said, you know, we don't really do that.
But it definitely seems like it's pretty prevalent.
And we haven't done any mandates inside of Airbnb.
It's all been organic adoption.
And I'm really, I'm excited to say a lot of people have said that adoption seems pretty high.
We actually, we want to share some numbers.
97%.
of active engineers inside of Airbnb are using agentic AI on a weekly basis so far.
90% daily.
And this is like agentic AI.
So it's not just chat.
It's nothing like that.
It's cloud code.
It's codex.
They're using agentic AI.
It's almost totally ubiquitous.
And this adoption has been like nothing I've ever seen.
It's really happened just, I mean, these tools came out a year ago.
So really, we've gone from basically 0% to 97% adoption in less than a year.
And I want to sort of share a little bit about what we did to sort of encourage this.
I think I was at the Netflix talk earlier.
It was very similar.
We saw a similar playbook to what you guys have been doing.
But the first thing we did was, you know, you ask a product guy, I'm going to say it was all about the product.
Like we launched an internal product called AirChat.
And we built a brand around it.
And we treated it like a product.
And AirChat was our...
It's like our internal AI harness, before that was a term.
And it makes AI work really, really well inside of Airbnb, like right out of the box.
So we handle authentication and permissioning.
We load default MCPs, integrates that knows how to operate inside of Airbnb.
You know, we created t-shirts, we ran workshops.
And this became kind of a rallying call sort of internally.
Everyone knows AirChat.
It makes it easy to share and talk about.
And it really helped kind of the word of mouth and positivity around it.
The other part is that...
We've kind of been doing this for a while.
So, you know, we've been on this AI journey for kind of like three years now.
So like many of you, we rolled out some of the initial like enterprise vendor solutions, like the autocomplete stuff.
We were pretty quickly realizing that that wasn't going to get us where we wanted to go.
We didn't have access to the latest models.
It didn't work across all the IDEs that we use.
So that's when we first launched AirChat as a series of IDE plugins in like JetBrains and Xcode, backed it up by the latest models.
got the ball rolling.
And then like many of you, we did things like helping like assist and speed up AI migrations.
We launched Agentec AI pretty early last year when that first started rolling out.
And so basically we've had kind of this long runway of people, you know, people have talked about before, but like, you know, you get on the learning curve, right?
We've had a lot of time for people to adopt, to learn, to sort of embrace these tools.
And we've been pretty, pretty lucky to be early at kind of each new wave.
The other thing is that we made it work really, really well inside of Airbnb.
So we really spent a lot of time building out our plugin ecosystem and marketplace.
This is an internal browser that I built myself.
I vibe coded it in like three days using AirChat.
I haven't been an active coder in like 10 years.
And in the old days, I would have had to write a PRD and get this on the engineering roadmap, but it would take like three months.
And I did it in three days.
But you can see like, you know, we blurred it out, but essentially this connects to everything that we use internally.
And this is all community managed.
And so it's a combination of, you know, there's some managed centralized tasks by different organizations, but then there's a lot of experimentation.
There's a lot of skills, a lot of, you know, instruction markdown files that people are using.
And we really encourage that as well as we all learn.
The last piece that I think has been even the most critical to all of this has been the peer-to-peer learning and the community enablement.
So, you know, I went back and looked up.
uh you know the very beginning of our ai for dev productivity channel uh and we we launched it about three years ago asking about some slides for a deck around how we're using ai in august of 2023 so god knows what was in that deck um but nowadays uh we never asked everybody to join this it wasn't ordered or anything like that and it's essentially um you know, basically every developer inside of Airbnb.
I think it's like one of, if not the most active Slack channel.
It's outgrown the AI team itself.
People are like contributing and sharing knowledge.
And this is just one example.
We've done a lot of events like hackathons and learnings.
We also have an AI champions program.
We did a train the trainers program where we taught AI champions to run their own workshops.
And this has really helped us sort of get to that 97%.
All right.
Myth number two.
Very topical.
AI has minimal impact on productivity.
So I'm going to share some numbers.
You saw the news, like, Justin, I think you published this a month ago.
You heard it this morning.
Productivity gains are around 10%, 7%, 8%.
And the main argument that you hear around this, which sort of, I guess, kind of makes sense, is that, well, engineers don't actually spend that much time coding.
You see various estimates from 25%, 20%.
I think the number this morning from Microsoft was 14%.
I'm not going to argue with the data, but I think this is about to change in a big way.
Saying engineers only code 20% of the time, it's sort of like saying golfers only hit the ball 20% of the time.
They're walking the course.
What are they doing?
They're not always hitting the ball, but it's kind of the whole point.
But a lot of what a golfer will do is like, you know, they pick their club.
They like line up the hole.
They check the wind.
You know, they line everything up.
And why do they do that?
They do that because they have to get it right.
You know what I mean?
It's like a high stakes motion.
So I think all these meetings, coding is kind of similar.
Like we have all these meetings.
We have all these reviews.
We have all these docs.
And I think a lot of it is about getting the code right.
You know what I mean?
Because it's expensive to change.
It's an expensive, like time consuming operation.
But that was how it used to be.
Right.
Coding now is cheap and fast and easy.
And what we're seeing is that time spent coding is going way, way, way up.
So I think, you know, and if you're not, I think you need to lean into this.
Like if you're still in meetings all day and you're still having people code 15 percent of the time, you're leaving like a ton of productivity gains on the on the table.
So this is a metric that we track internally that looks at how many hours per week they have an active agentic AI session.
And this number has been going up and up and up.
And this is now.
This is much higher than 20%.
And this is the median developer.
If you look at the top 10% developers, they have an active agentic AI session running more than there are hours in the week.
Because coding used to be about hands-on keyboard time.
It used to take a long time to get into the flow.
It was hard to do small bite-sized changes.
All of that has changed.
You can be in a meeting and actually be coding in the background, right?
Because your agent is using it.
So I think what we're seeing is that people are actually spending more and more time producing code.
So what does that mean for PR throughput?
The big number, the hot topic.
Our PR throughput inside of Airbnb is 65% higher than it was before we introduced Agented AI.
So this is a lot more than the 7% or 10% that you may see.
And if you look at the growth, it started growing a little bit last year, right?
And so we were also in that kind of 10% to 20% number.
But starting at the beginning of this year, this has really, really gone just through the roof.
And if you look at this, this is, I guess, what someone who knows more than me would call like a longitudinal study, I guess.
But like these same developers were, you know, you know, had kind of like industry average output a year and a half ago.
And now they're sort of kind of closer to the top band.
And we actually, we went a step further.
We drilled into this a little bit to see, okay, like this is going up.
Is this actually being driven by AI?
And if you break it down, so, so many people are using AI instead of Airbnb that, um, We had to go, like daily usage wasn't granular enough.
So we asked people how many hours per day they're using agentic AI.
And we broke it down to like one to four, four to six, six and more.
And what we found is that the sweet spot seems to be like around four hours a day.
People that were using agentic AI for four hours or more per day have seen PR output more than double from before AI.
And like the four hour day people also are seeing growth, but not nearly to the same degree.
And again, if you look at the more than four hour people, these groups were like identical two years ago.
It's not that these were more prolific developers before either.
These were basically had the same output.
And the only thing that's changed is how much they're using AI.
And again, like this is really just kind of exploded.
You can see like these developers went from kind of industry average output to like off the charts.
I don't know, 1%, top 0.1%, pretty much just from agentic AI usage.
Another way we look at this is AI authored code, another great DX study.
So again, this is also another DX number that they published.
For the industry average is around 27%, I think, as of Q1 for AI authored code.
Inside of Airbnb, we're about, we're double this.
So we're closer to 59% today.
So more than half of code inside of Airbnb today is primarily authored by AI.
And they tell us that, you know, they do these custom benchmark groups.
There's like 20 pure companies.
This is, I guess, a bit of an outlier where we're sort of at the top of this group.
Lastly, time savings.
So in terms of time savings, developers say they save around six hours a week due to 8i coding tools, basically.
Again, this is kind of also number one.
And frankly, this is undercounting it.
because the question was sort of broken.
We broke the question.
The top option was eight hours or more per week.
And that was the number one response.
So like over a third of developers told us they were saving at least a day a week because of AI, which was the biggest group.
So we don't actually really know kind of what the upper bound is.
And I think, you know, my prediction is that like at the next one of these events, like these numbers will look quaint for kind of everybody in this room.
Like if you're not, I think a lot of you are already seeing this, but The tools are getting so much better.
People are getting so much better at using them.
I think that we're like every estimate we have made on terms of like AI usage or AI growth has been like wrong.
We've undercounted.
And this doesn't really seem to be changing.
New models are coming.
New harnesses are coming.
New capabilities are coming.
People are getting better at this stuff.
I think that this is sort of going to come and come very, very fast for a lot of us.
All right.
So I'm like 20 slides in.
I haven't even gotten to the topic of the talk.
I apologize.
So I want to talk about the other one is like AI coding tools are just for coders, right?
They're called Cloud Code.
It's called Codex, right?
It's for coders.
I think, you know, if anyone who's played around with these tools realizes that they're much, much more capable of that.
So something really funny happened that we weren't expecting.
We thought our addressable market for these tools was our developers.
So we have around, you know, 2000 or so developers inside of Airbnb.
And as our AI tool usage went up and up, we're like, okay, we're going to get to about the size of our active developers.
That's our TAM.
And then it's going to sort of flatten out.
And it didn't.
It just kept going.
So as of now, we have about double the amount of users of agentic AI tools as we do active developers.
And we were not really expecting this.
What happened?
Well, turns out a lot of folks had the chance to go home over Christmas break, finally have a chance to play around with these models and said, oh, snap, these got really good.
When did that happen?
The last time I tried this, you know, it had, you know, how many R's were in strawberry or whatever that whole thing was, right?
And everyone was like, wow, this got really, really good.
So everyone came back to work essentially like just demanding to use them in their own day-to-day work.
And the way they're using them is like for everything.
And what really surprised us was that they're actually doing this like in the terminal.
So they just started using the terminal.
Our finance team, the director of product and finance, put together, I swear, like a 30-page onboarding guide to VS Code.
So he has onboarded like the whole ops and finance team to VS Code where he like has AirChat running in the sidebar.
He has like a file browser with Markdown files.
Like this was not in our PRD for AirChat with like finance team will learn to use VS Code.
Like that was not going to like that was not happening.
But that's what they're doing.
And they're using it for everything.
They're using it for prototyping, for data research, as a personal assistant.
They're building all sorts of really cool tools.
And we did one thing that turned out to be really powerful that Madison will talk about is like we built an SDK around this stuff.
So we built an SDK for AirChat so that anyone internally can build their own AI-powered applications.
And people are going like really deep on this and building things we never would have thought of before.
So just some examples like Pascal is an internal product planning tool that the product managers and designers use to go from idea to concept brief that like...
does a ton of market research, looks at our roadmap, pulls in business objectives and puts it all together.
We have a prototyping tool for designers so they can rapidly vibe code prototypes to have higher fidelity concepts.
I mentioned the operation folks in VS Code.
The other one that's been really popular is data analysis.
The data platform team built this whole data query tool with guardrails that queries business data.
These have been really, really popular.
I think what's really...
kind of most exciting, I think, a little bit about this and the opportunity is, like, stepping back a little bit.
You know, if you look at, if you look at, like, every technology, right?
Like, every new technology, the first phase is just applying the new technology to the old way of working.
You know what I mean?
This is sometimes called, like, the horseless carriage effect.
Because, like, the very first automobiles were basically just carriages without the horse.
It was called horseless carriages.
You know, the first movies were, like, people just pointed cameras at stage plays, right?
And in tech, like...
You know, the internet came out.
The first websites, I mean, they were just like paper brochures.
You know, the first news, it was like the front page of a newspaper.
Or when mobile devices came out, like the first websites were like full HTML websites and you had to like scroll in and hit the dropdown menu.
And like, it took a while for people to figure out like the newsfeed and scrolling.
And I think that's sort of, I think that's what's really exciting about this is like, we're kind of trying, we're hoping to speed run that part a little bit.
We kind of want to rush ahead.
It's sort of, we look ahead, it's a bit of a fog.
It's like.
wheels on luggage.
It's like sort of obvious after the fact, but like, how did that take so long?
Right.
Um, so we're starting to rethink kind of from first principles on the product side of like, how do we do these tools?
How do we do what we used to do in like an AI first way?
Um, and so I think, um, they would have a prototype in earlier, and this is like a big one for us as well.
Um, in the old way of working.
You know, despite all of our lip service to like agile, like iterative development, it tends to be kind of waterfally, right?
Like a product manager writes a long requirements doc, hands it off to a designer who creates mocks, who hands it off to an edge to create a prototype.
They go back to the designer, the designers are like, oh, well, now that I see it live, that's not quite what I want.
And the product person was like, well, that's not really what I was envisioning.
And, you know, it kind of takes forever.
In an AI first world, We're kind of flipping this on the head in a couple ways.
One is, you know, same thing.
No more long requirement docs, right?
Let's just create a shorter concept brief that gets the key elements in place and go right to prototyping.
Prototyping looks a lot more similar to what eventually you're going to launch.
It's higher fidelity.
Let's just like start there.
You know, again, let's go crazy.
We don't have to build one prototype.
We can build five prototypes.
We can build 10.
It's cheap.
It's really easy.
The other thing is to be way more collaborative earlier on.
So instead of this sort of like PM is responsible for this designer, you know, and we really want to create smaller pods, like smaller teams moving faster is sort of our mantra and collaborating early on.
And then, of course, number three, using AI, right?
Leveraging AI to create the concept brief, create the prototype, and even going further, like things like, you know, we talk to these teams and that they often say, well, the hard part actually isn't creating these assets sometimes.
It's just.
keeping everything coordinated, right?
Because you create the prototype, you learn something, you update that, but the doc doesn't get updated.
And then when you go to review, nothing's in order.
We can use AI for that too.
You know, we can have AI project coordinator, create a Jira ticket, create everything in sync.
So this is kind of early days, but we're seeing a lot of real promise here.
And it's like something that I think is going to help transform.
you know, again, I think they talked about the North Star of like idea to value.
We have sort of a similar concept, right?
As coding becomes less of a bottleneck, it's going to be these other steps in the pipeline that become more and more critical.
So last one, maybe the most controversial, I don't know.
Vibe coding gets sort of a bad rap.
You know, vibe coding, it's not coding, it's AI slop, you know, yada, yada.
Yeah, no, like vibe coding is the future of coding.
Like this is how engineers are coding nowadays.
So I mentioned AI authored code.
So same group of people inside of every 28% of developers said that AI primarily authors 80 to 100% of their code already today, which is a pretty big number.
But the speed at which this has happened is kind of mind blowing.
So this was our last survey in October.
3% of people said that AI authored 80 to 100% of their code.
And you could, I mean, you can see the green, but this totally flipped.
Six months later, It's 28%.
Six months ago, more than half of people said AI authored 0% to 20% of their code.
Now that number is 11%.
More than half of people now say AI authors more than 60% of their code.
And again, these are the worst the tools will ever be.
Like, this is only accelerating.
Like, we're just at the beginning of this.
You know, Andrzej Karpathy sort of famously kind of summed this up earlier on.
He was like, this is the biggest change in programming, and it happened in a matter of weeks.
And I think...
if this hasn't happened already in your organization like it's coming very very quickly and soon um and i think you know there's a lot of concern right now about ai adoption ai productivity like yeah that is a concern but like skip a step because i think you're going to have a lot of problems very soon about overwhelming code review too many like you're gonna have a lot of other problems on the other parts of your system because i think this is coming whether you know you know no matter what you do um the other part of this is um You know, code quality, right?
We're all concerned about that.
Like, you know, we're moving faster, but is it coming at the sacrifice of quality?
And so far that hasn't really showed up for us.
Our change confidence in code maintainability scores are basically flat.
Our change failure rate has actually gone down over this time.
Code maintainability has actually improved as a sentiment.
I think largely because we've invested a lot in like migrations and tech debt.
So I think our code base is healthier.
So this is something we're watching closely, but so far we haven't seen this.
kind of concern or trade-off around quality, but we're also actively working to maintain that as well.
I think what's really happening is that, if anything, we're not actually fully unlocking the potential of this because of the tooling.
The tooling is actually holding us back.
It's not even the quality of the models or the harnesses at this point, but there's some real challenges around the tooling itself to kind of unlock this next phase.
So, you know, we talk about this as well as like asynchronous AI.
So, you know, we started with chat, right, where you ask a question and get an answer.
We're in kind of the agentic phase now where you ask an agent to do something for you.
But we're moving very, very quickly to this world of async AI where you're managing multiple agents.
We're already sort of seeing it with power users.
But there are a few kind of key gaps, I think, right now in the tool chain.
One is that it's still really, really hard to manage multiple sessions, right?
Like if you're in the terminal, you have a couple tabs open.
It's kind of a nightmare.
You don't really know when you need to jump back in or it needs your attention.
There's still not a great user experience about this.
It's also really hard to run sessions remotely still.
I mean, there's some good recent Feaster launches, but I think this is largely unsolved.
You know, you don't want to have everything tied to your laptop.
You want to shut it down.
You want to go to bed.
You want to get on a plane.
You want to pick it up for your phone.
You really want to be able to run these things remotely with all the right permissions and security in place.
And then lastly, chat, I think, is underrated as a UX.
It's been pretty good, but it isn't always the best, right?
Like you run an agentic session.
You have things running in parallel.
You jump back in and you're like, okay, like what was going on?
Like, what was I talking about?
A lot of times the operations that engineers do are pretty standard, right?
It's like rebase and merge this branch, like open a PR based on these comments with these fixes and like submit a fix, right?
Like this doesn't need to be a chat.
Like you can just have a button.
So we're working on something internally called AirChat Remote, which is sort of our next generation version of AirChat that addresses a lot of these questions.
So.
Right out of the gate, workloads run on something internally called AirDev Workspaces.
And Madison will give you the real scoop on all this stuff.
So that runs off of the laptops.
There's a whole UI built around managing multiple sessions in parallel.
So you can very quickly and easily see each session, what the status is, what kind of tools are running, which ones need attention.
So you can kind of jump in and out.
And then lastly is that...
we do have these sort of dynamically created shortcuts, right?
So if we have a pretty good idea of what you need to do next to keep this thing moving, we're just going to expose that as a button.
All of these are backed by full agentic sessions you can jump in and out of.
But the idea here is that, you know, we're really sort of leaning into what's really possible.
And this is live.
So this is an EAP internally.
We have like 50 users.
And I like, I want to scare people, but like the PR throughput for these folks is like 3x higher again.
So like, I think the wave is coming and it feels sometimes like we're behind, but I don't, I don't think we're going to feel that way for very long.
So just to close up, I want to sort of think about this as AI truths, right?
Like AI is all about augmenting humans and unlocking creative potential that I think is pretty unprecedented.
The one word we hear the most is, is, is fun.
People are having like more fun than they've ever had before.
You can get to high adoption just through organic means, right?
And honestly, that's what you want.
Like you want people voting with your feet.
You want missionaries, not mercenaries.
We're starting to see really significant impact on productivity from not just engineers, but non-engineers as well.
And more and more, I think this is going to look like, you know, a step removed from the code itself where you're managing async teams dynamically.
And I'll hand it off to Madison, who's going to go talk more about that.
Thanks, Christopher.
All right.
So how did we, sorry, too long not talking.
Okay, how did we get here?
I won't have time to dive really deeply into too much of the technical detail, but I do want to go over some of the goals, some of the requirements we set for ourselves, as well as a little bit of the architecture so you'll get a high level sense.
I do want to start with our overall philosophy, though.
One thing that we've focused on this entire time is not starting from a particular tool that we want to bring in.
It's more trying to predict how developers are going to work in the future, in some near point in the future, as far as we can predict out, seeing what we can reuse, seeing what we have to invest in ourselves to build.
And then as the ecosystem changes, we do this exercise over again.
We set a new North Star vision that we're heading towards, and we continue to do that.
This has left us with a few ways of working that seem like they might be in opposition a little bit, but that tension actually helps us balance for short-term and long-term development at the same time.
We are trying to be about four to six months ahead of what you can get from the industry out of the box.
And so if we can build something that will provide value in that time with a reasonable lift, we go ahead and do it.
That does mean that we have to be open to swapping out pieces with external solutions as they come along.
And so we try to build the fundamentals in a very modular way so we can do this.
That means we don't always have the fanciest or shiniest things immediately when they're available.
But that does also mean that we're trying to minimize churn on our paved path, right?
So what we build, we make sure it's good, it's solid, and we tell people this is the way that you should do things.
And we don't change that on them too often.
That doesn't mean that we don't like innovation, though.
So we really try to embrace community innovation and developers and teams unblocking themselves.
Seeing the ways that people actually go about unblocking themselves is a very helpful signal for us in figuring out what our paved path should actually look like.
And so often these innovators or first requesters become our really close partners in actually developing the paved path moving forward based on some of these things they tried for themselves.
So unless there's a security implication or a cost implication, we try to let people feel free to experiment.
Christopher's already talked about this evolution, but I do want to zoom in to when we started to bring agentic AI to Airbnb because that's the point in time when AirChat went from being an IDE plugin into really being more of an ecosystem.
So one of our goals after we did a bunch of third-party evaluations was really to find a way to provide unified capabilities across all our surfaces.
Our developers really didn't want to give up their IDEs.
They loved it.
We wanted to meet them where they were at.
So we introduced something that we called AirChat CLI.
This is a light abstraction wrapper around coding agents and passes all of the APIs through a single unified gateway.
So we get cost control.
We can see kind of usage metrics in a very standardized way.
But more than that, it allowed us to call into it from our various IDEs, as well as provide this new CLI option for people to do coding.
One of the things that, as Christopher already mentioned, people really like to do with agentic AI when it first came out was look at migrations.
Migrations are very toilsome.
For small or relatively simple migrations, this was pretty straightforward.
You could just use a coding agent, handle it all yourself.
But when you start getting into more large or complex migrations, we found that a structured approach actually worked a lot better in order to help manage the unpredictability of AI.
So we developed a framework that allowed people to use a sequence of steps on each target where you would validate, transform, validate again, and do that in a loop until that particular step passed.
Now, this has worked really well.
We have dozens of migrations that have happened.
We save on average between a few months or up to even seven years per migration.
It's about a 5x speedup, which is great.
However, and as Christopher mentioned, people want to be able to close their laptops, walk away.
They don't want to babysit this.
We didn't really have a great solution for them at this time.
We'll come back to that.
All right, moving on, we've gotten into what we're calling the AirChat SDK.
The goal, as we quickly saw, as individuals became comfortable with agentic AI for their own use cases was, how do I empower other people to be able to operate better in my domain of expertise?
And so people really wanted to build tools that they could share with other developers to help them work faster or better in their particular domains.
You can enable sharing of MCPs or skills or plugins, etc.
But they wanted more than that.
They actually wanted the ability to create their own customized UIs on top of this that would live in a web interface.
So we introduced the SDK.
A couple of technical requirements for this.
We really wanted a language agnostic protocol for communication and data modeling so people could use this in whatever language they were comfortable.
We wanted bi-directional communication also so that we could have this true back and forth conversation with the agent, even from the web application.
And we really wanted people to be able to configure prompts, tools, hooks, anything that people needed to customize for the particular application is something that should be able to happen through the client.
DataCo, as I already showed, this is one of our data applications that people have built as high hundreds of daily active users, which is fantastic.
It's quickly becoming one of the primary ways to actually look at data.
You don't have to know what tables you're looking for.
You don't have to know where the data lives.
It's probably the most successful application so far that's been built on top of our AirChat SDK.
But we have about 40 of these projects that are now stood up internally.
It's pretty fantastic because we've taken the time that it takes to build one of these from a day or two down to about two to four hours.
So a lot of teams are starting to feel really empowered to create things that actually fit their individual needs.
All right, this takes us to today when we start talking about parallelized and async AI.
So we have a bunch of goals now.
People have gotten very comfortable using agentic AI and they have new needs.
We have people that are doing parallelization on their own.
They're opening multiple terminals or trying to set up get work trees.
It's really cumbersome for them, and they never know which ones actually need their attention at any given moment.
People are also coming up with all of these various use cases for where they might actually call agentic workflows.
They're things like triggering from an on-call support channel in order to have some agentic triaging happening and actually fixing things, fixing things on PRs when they fail, or really just starting workflows from Slack and then being able to check on it later.
So there's lots of interfaces people would like.
There's also a desire to have these long-running non-local environments that persist context, particularly referring back to the migration case.
This is one of the main goals that we are hoping to solve.
People can run these things, they can walk away, they can come back, they haven't lost their place.
Which takes us to this grand goal of trying to truly offload work from developers.
What things can we run asynchronously in the background and autonomously?
Turns out that actually solving all three of those previous goals is needed for solving this as well.
So as Christopher mentioned, we're really leaning into the AirChat remote idea.
This is really just a starting point, web UI.
The idea of this is supposed to be that you can see everything.
And for some nomenclature, we're calling parallelized sessions sessions, and we're calling sort of asynchronous workflows tasks.
But you can manage both of them through the same web UI.
But the magic is really probably in the architecture.
So we've designed this.
so that it connects to and sits on top of all of the previous pieces of the AirChat ecosystem, which is why I ran you through all of those.
We hope that one day there's going to be more vendor solutions that we can kind of slot into pieces here so we're not having to maintain this forever.
But in the short term, this really provides us a powerful way to be able to support multiple agents, to be able to have it not run on people's laptops directly, and to do sort of parallelized synchronous workloads as well as the asynchronous ones.
AirChat API.
is kind of the central coordination point of all of this.
It's where all of the client interactions come from.
You can do web UI, you can do Slack, you can do from the SDK apps, which is going to be pretty fantastic.
It handles the authentication, the lifecycle management, and it streams events back to the clients.
It takes us to our worker pool.
Our worker pool is built on top of something called the AirDev platform.
So we have infrastructure for developers at Airbnb to allow them to spin up their own customized workspaces or cloud IDEs.
in a way that is remote.
They're pre-configured for fast development on our major monorepos.
So out of the box, they work very nicely.
So our strategy today is to re-leverage these cloud environments, not only for parallelized coding agents, because they're really easy to configure and stand up, but also for asynchogenic flows, because they do provide sandboxing.
Even though each workplace has access to our monoreable toolchain, build dependencies, source control, everything an agent might need to be successful, they are network isolated from production services and user data.
Last but not least, we've got the agent, Damon.
He runs one per worker.
It's really responsible for configuring the agent tool loop of a task on a workspace and starting its execution.
It wraps agent runtimes behind ACP.
For interactive sessions, we go back through the remote API so we can really stream the responses back to clients without needing daemon connections directly.
Christopher mentioned this is still in alpha.
We're still doing a lot of additions to this.
I think right now as I'm speaking, someone is adding in the ability to run these things on a schedule.
So that UI should exist by the time I get back to the office, which is pretty phenomenal because somebody put up a design for it this morning.
So we can build faster than we can align, I think, sometimes.
We also want to build a migration platform in here to make it easier to use our migration framework directly, since we know this is one of the primary points that people will be able to find value from it.
as well as handle sort of more non-deterministic migrations.
So big picture, what's next?
Magic 8-Ball would probably say, response fuzzy, ask again later.
This is due to us projecting our vision out into the future.
We're about at the point where we need to do that exercise again.
But we do have a few predictions.
We think that AI-first architecture is going to become increasingly important.
We do think that there's going to be increasing importance also on guardrails as code goes through.
Throughput is going to accelerate.
The more we can do in automatedly checking that this is great is going to help us reduce bottlenecks.
And we also think that there's an increasing importance in scale and resilience of the surrounding infrastructure.
I think, as someone mentioned earlier this morning, the more goes into the front of the pipe, the more needs to come out the back of the pipe.
The whole pipe needs to expand to accommodate that.
So as for specifics, we're still figuring that out.
If you're working on something similar, if you have a vision of what's next, we'd love to chat with you.
But that is it for us.