Hey y'all. I'm Quinn from the AMP team. We'll make this super interactive, so feel free to yell and shout. I am very used to a large volume of people yelling and shouting at me. I got three kids and we have a shared Slack channel with the Tempo and Paradigm teams and they're very demanding about everything. I want to run you through how we build AMP. We are, you know, one of the big agents and all, but let's start out by—I heard you all want some free credits for AMP. So, I'll kick something off while we go. Anyone else use dictation a lot when they're— Yeah, dictation is the best. And I think we can do the ultra fast. I'll say make it so if you make a new thread in AMP and it has the string 'frontiers 2026' in the user message, then it shows and it'll transcribe something with like a cool animation like the Frontiers Paradigm style, and you can click it and you can get like 50 bucks. I made that lead AMP frontiers offer code. So you can link people to that. And also, I am presenting to this whole group in the AMP design talk thread. So as it's like shipped to main and deployed and all, can you go and show a little thing at the bottom of my slides that I'm presenting? It's in the AMP design talk thread and yeah, I don't think I need to review this, so you can just push it to main.
All right, we'll let that go and we'll get some updates as we go. I also made it look like this. Does this give anyone a headache? Is this— Yeah, I don't know. We'll see what looks good. All right. So, AMP, we are an agent. For those of you who have not seen or heard of AMP, it's something that you can use instead of Claude Code or Codex or any of those other things. And yes, you know, everyone is building the same thing. We want to stand out. This is the most competitive space I've ever seen. And also for all of you, if you're using an agent, it's a really big choice for you, your company, whatever. Of course, what everyone does, they just pick seven of them. And I'll talk about why I don't think that's a good decision.
So, we've tried to go as extreme as possible. And AMP is all about finding the extremes. If we fast forward the way things are going, how would we build our company? How would we build our product? And the obvious thing is, ever since AMP started, is, well, of course, we're going to build with AI. And now this doesn't seem that controversial. Although there's still people out there who say, oh, anything that's built with AI is total slop. But then we've built something that people think is pretty delightful. And we want to be a shining example of how you can build something that is delightful, even if it is built with AI. AI does not mean slop.
And we've also made it so the company is like how we imagine all the companies of the future will be. So we are one of the big agents. We have 21 people on our team. And you know, I say that I think probably a lot of you are smaller companies, so maybe that sounds pretty big to you. But all the other guys, all of the other people, companies building agents, they're like 500 people, 1500 people. I saw one of our competitors say that they hired 19 people and they started on Monday, and that's our entire company.
It's crazy. And I don't understand how you have the right back pressure to build a really good agent, to figure out what it means for your company for everything you do to be through the agent, if you're hiring a marketing team and that marketing team has their own marketing team. I think that's crazy. So we're trying to live exactly how we think the future will be, and we've seen the alternative. I was the founder and CEO of Sourcegraph. You know, big SaaS company, like 200, 300 people. This is a lot more fun, I will certainly say.
All right, I'm going to walk through how we are using AMP to build AMP, and I'll share some of the alpha, some of the things that we found really helpful. None of this is AMP specific. I think that you could probably figure out ways to do these things in other agents or, you know, they'll copy what we have. But just to give a lay of the land, you know, what is AMP? It's a TypeScript codebase. We've chosen all TypeScript except for some iOS stuff. It's a lot of code. We don't really look or care how much code it is, but we pulled the numbers here. And we deploy constantly. So, as GK said, we do not do any code review. We've never done any code review. We think it's the dumbest thing ever. And the one thing I disagree with is he showed that big exponential about PRs going up. I actually think that they're about to go way down because pull requests slow you down. I mean, we are shipping 90 times a day, and when we push something, it takes seven minutes to get to production to all of our users. We don't do like self-hosted. We serve big banks and all. If we were to do code review, that seven minutes somehow turns into, I don't know, a day, and it's like the worst part of any human's day to do code review, especially when it's agent code.
So this is like the lay of the land. This is what I'm talking about our application. And let's walk through some of the things we do with the agent that we found very helpful. And a lot of these are techniques that would not have been possible one or two months ago.
So first thing is you should never be waiting on the agents. You should be giving them really long leashes. If you are actually caring that much about how fast the agent is—unless, like I did ultra fast because I'm giving a talk to all of you—if you're actually caring, then you are probably not giving them a long enough leash. You're not telling them enough verification steps that they should do. You're only using them for a small part of the process. You should widen that. Have them help you with figuring out what to build and how to build it, not just the building. And then once they've built it, have them actually ship it and follow through to production and make sure that it fixes the issues. You want to give them a much longer leash and really challenge yourself to do that. One of the hardest things about building an agent and using an agent is, you know, like GK said again, you got to like put your mind in a state of openness. And I suspect that you all are way better at doing that than most people out there. But you like, you got to constantly be thinking, how could I have done that better? And it's honestly tiring sometimes, but I don't think we all would have it any other way.
One tactic you can use here is never iterate with the agent. Never say like, 'Oh, do this, change this, change that.' Because then every single time you're going to wait for the agent and then it's going to come back to you. By the time the agent comes back to you, you want it to give you like the whole menu of options. And this is great. In theory, you could do this with a human, but if you ask a human to go and spend, you know, three or four days building a bunch of things knowing that you would discard all but one of them, that person's probably going to quit. Like, if they're good, you know, if they're not good, you wouldn't want them anyway. But you can have agents go and show you a menu of options and you can say, 'Yeah, you know, let's pick C.' So, I use this to make that little waveform when you're dictating in AMP. And it's these little like delightful details. I really wanted to get that right and I had no idea and I just had to come up with a bunch of options. So, have it go and do that work.
One other thing that's really helpful is thinking about what would be irrefutable proof. So, okay, you know, thought experiment. Let's say your agent did a bunch of work and like the diff was pretty big. You hadn't looked at it, and let's say you know there's something that said like there is a billion dollar bond that the agent is posting and it guarantees that it is correct and it's also, you know, backed by, you know, GK and Jamie Dimon and all that. Like, yeah, you'd probably feel pretty comfortable clicking because if there was a bug then you'd get that billion dollars. Obviously, that's like a thought experiment. That is not actually what we are doing or anyone else is doing. But there is a possible way that the agent could make you 100% comfortable with shipping and put the onus on it. So, you know, one example from one of my threads is I wanted to make sure, you know, there's this like icon jiggling in AMP. And you know, if it just shows a screenshot, I don't know if it's actually fixed that. I had it look at every single frame and prove to me that in every single frame it did not jiggle. So, you know, my kind of chicken scratch prompts. Prompts are, you know, there's no tricks to prompts. You just got to say what you want. And then it went and analyzed the video every single frame and proved to me that it didn't move. And I can kind of look at this and, you know, see that it's correct. So, ask for irrefutable proof. And if you don't know what irrefutable proof is, well, the good news is the agent is pretty good at figuring out what it is.
Many eval. So, you know, we think that a lot of these models are trained at being AI researchers by the labs. So, it's no surprise that they're pretty good at running little evals. And if there's something, especially in like a non-deterministic part of your system where you wanted to test, you can just have the agent run a bunch of tests and it knows all the right ways to do that to isolate it. They're just like really good at it. Any of the models really. So whenever we make changes to our own system prompts or tools or instructions, we just have AMP go and see if you take the local dev app and you run, you know, three times or 10 times—you don't need like thousands—does it actually work? Here's like a dictated prompt. As you can see, it's pretty messy. And yeah, and you know, I was looking for it to be zeros here in the column. So, it's also, you know, really helpful when building any kind of agent thing because these are all so non-deterministic and you want to run them and see how different models respond. You could also expand this other dimensions. So, you know, this I think just tried it on one model, but if you got to support a lot of different models, then have it do, you know, testing on a lot of different models.
All right. So, me specifically, as I said, I use dictation constantly. It is, you know, ultimately like using an agent is about getting what you want to the agent. I know there's some people that say like, 'Oh, I have these like 300 skills or like this trick.' Like, there's no tricks. Like, I think tricks are dumb. It's all about what you want and getting it to the agent. And if you are typing, it feels like breathing through a straw. You got to get in the mode of just dictating. If you're working in an office and you can't dictate, well, that's a good excuse to not work in an office. Or, you know, one time on an airplane, I had like a little microphone and, you know, the flight attendant came up to me and there were like people complaining and worried that I was, you know, doing weird stuff. But find a way to dictate, no matter what. You can see, you know, I struggle with this too. We had a meetup in New York and it's a lot harder to actually dictate there, but you want to be dictating.
I am only using the latest models. Like GK said, you want to put yourself in a place where you have nearly unlimited token spend. And as a company, if you don't—if you're paying someone a lot of money, a human, and you don't give them a lot of token spend, I mean, it's like the dumbest thing ever. It's like if you're like an automotive worker at Ford, you know, motor at the factory, and it's like if you make those workers pay for their own screws and like power tools, it makes no sense whatsoever. There's models that certainly are cheaper if you take a known task. Yes, you can show that it's cheaper, but 10% of the time it's going to get it wrong and you don't know which. And really the vast majority of tokens, the vast majority of spend, and the vast majority of things that people create with AI are created using the latest models. Most people here certainly are in competitive fields. So if you're in a competitive field, your competitors are probably using—the ones that are winning are using the better models and you have to be too. If you're just like playing around, yeah, you know, you can try out other models, but if you're playing around, you kind of want to have the most fun with the latest models.
I think we've seen a huge migration. People getting away from their computer, people getting like iPads and phones and even the watches. I think you want to find a way. So, there's a way to get information from your brain to your agent no matter where you are. So, we've got, you know, that— Oh, yeah. If you haven't seen that, that's Puck, our kind of mascot avatar. So, anywhere you are, you want to be able to send stuff to your agent. And it kind of sucks if you've got like your work coding agent and then your personal agent and you know on the go you can only do it to your personal agent. I think you want—you know, like really the walls are breaking down between them and you want to be able to get your intent anywhere.
That zero friction is really important. It might seem like agents have low friction, but if you got to keep your laptop open, if you have to be, you know, just in the right app, that friction is really meaningful and it's going to prevent you from using the agent as much as you should.
Here, let's check and see how this one is going. All right, this is where I was making so you get a little offer. So, it's working and— All right, here I'll make that a little bit bigger. All right, back here.
It's crazy to think of how the usage has changed in agents. I think a lot of people here—who here at the beginning of the year was using an agent in their editor? Yeah, don't feel embarrassed to admit it. And CLI as well. We killed our editor extension and we haven't killed our CLI, but we're not using it anymore.
Like the workflow has already changed twice this year. Basically, you know, to a first approximation, all agents are running in the cloud now because when you have the ability to do that, then not only can you do so many more things, but you can have the agent do things like the mini evals where it can spin up its own work and it's not tied to your laptop. I've never seen any kind of product where like you have to reinvent yourself and reinvent the product so often. But yeah, we are entirely in the cloud now. We call them orbs and other people call them cloud agents and doing stuff from, you know, native apps and from the web, from the watch and so on.
All right, that's how we're using the agent. I want to talk about how we as a company are working. And this is really just following the principles of the agents are really good at the building. They're getting good at more things. And what is the company of the future going to look like? We have no roadmap. We work with a lot of other like companies, like infrastructure companies and so on, and it's so dumb when we like need a feature from them and they say, 'Oh yeah, you know, that I'll talk to my PM and we'll see if we can get that on the roadmap and it's Q3.' And then there's other companies where they say, 'Oh yeah, I'll ship that tonight.' And that first kind of company is just not going to make it. And we certainly don't want to be that. We have one meeting a week and we don't really have any code ownership. This is another big blocker. If you've got a really great agent, let's say your company is super supportive, but anytime it strays outside of your little area, if you have to go and get permission from those teams, there's nothing that drains your energy more than that. And there's nothing that constrains the agent more than that. I think that kind of like how Google had this principle of they had one big source tree and for your little bit you had to make the tests so good that there could be like a, you know, dumb evil person who would still be unable to break it. That's how you should think about your codebase with your agent. You don't own anything, but you should make it so if you work on something then a future agent will be able to see how to not make mistakes basically. And a lot of that is how can you get your human intent about why you did this or, you know, why this is that way into the code, into somewhere near the code so that a future agent can find it and then means your team can go fix stuff and all. And that's part of how we're a 21-person team because anyone can do anything. You want to get to that kind of state.
And then this is the most woo-woo ever, but we always say work on what gives you energy. That is certainly not how things were in like 2022 tech companies. It was like work on whatever the salesperson tells the PM, tells the project manager, tells your engineering manager, and tells you. And that kind of sucks. So work on what gives you energy because one person with a great idea can do so much more than ever before.
Our velocity has gone way up. We've been using an agent, you know, ever since agents like became good. But with the move to cloud agents, our orbs, we've seen that go way up. And with the models getting better, with us being able to give them much longer leashes, our velocity has just gone like way, way up. There was a period where we had to completely rebuild AMP. So, anyone remember this period from like March till April where AMP like didn't really do anything? It was kind of slow and lame. We heard that from the Tempo folks. Yeah. So, I'm sorry about that. But, we got to the other side and we basically completely rebuilt the entire product. And you know, you don't usually do that in software, but that's what this kind of calls for.
Feature flags, we use them constantly. They've always been around, but they were always a pain. You know, I remember there's companies where they would have these old feature flags littered in their codebase and it would add so much complexity and if you wanted to add a feature flag, you'd have to figure out, you know, what are all the places you add it. Now, the agents are really good at it. And what we do is when we have a new feature that we think is cool, we'll ship it and we'll let users go and, you know, find this, they can toggle it on. And then we've given the agent read-only access to a lot of tables in our production database and it can get access to anything if it has like a, you know, three-hour temporary grant that we have to approve, but by default it can access anything and it can see everyone who's toggled that on and how has that changed their usage. Have they been successful with that feature? Have they done more things? Have they hit errors? Because it has access to our G-Cloud logs as well. We can see if they toggle it on and then toggle it off. We can email users and ask them for feedback. The agent can then see what they say in response and give them $10 in free credits. And there's this whole workflow around feature flags and letting people opt in that because the models are so good at writing code that is gated by feature flags and they can just do that on their own, then it enables all this other stuff essentially like self-service and autonomous.
So when we're done with a thread, when it's pushed, we let it stay open and we let it keep, you know, monitoring for changes. So, you know, here's an example. We added support for a lot more BYO options in AMP. This is the kind of thing that's like painful. Every BYO API endpoint is different and huge potential for bugs. But I would just, you know, say out of all the users that want this, you know, which 10 users should get it. Then when we open it up, you know, I would say like for these users that have it, what have they used? What providers are they using? Have they hit issues? How much are they paying? All that stuff.
And without this, you know, this is the kind of thing where you'd have to go hire a PM and probably a lot more engineers and that's going to slow everything down.
All right. Yes, we do not do any pull requests. Who here has like reviewed with your feeble human brain? Okay, that's leading the way. Let me ask you— Who here has done a PR review in the last week? Okay, cool. By feeble human brain, I just meant that in comparison to any model out there. No judgment against all of you.
Who here thinks it's like the dumbest thing ever and that it's just like cargo culting what people used to do? Yeah.
Does anyone get joy from reviewing code? Okay, there's one person. Thank you. You can review all of our code.
And someone asked about how do you not do code review if you're dealing with financial code? What we're seeing, you know, at least in big banks—I don't know if I can say the word 'big banks' in here without like having stuff thrown at me. But, you know, let me— Yes.
That it's software companies that think that they're hot because they got good at writing software and that doesn't matter anymore. They're the ones who are clinging on to their dumb old software development processes. The big banks, they're not sentimental about this. They just want to go make money. They don't care about code review. They just need to make sure that can they get their regulators to go along with it and does it produce a good product in the end. You know, any other company that's not like a software company that was, you know, big in 2022, they don't give a damn about code review. It's just these companies that are clinging on and everyone knows it's dumb. And if there's anything that you know you can take away from this, go back to your company and say, 'Let's start not doing code review on 20% of the PRs. Let's start risk scoring them and then, you know, grow that over time.' What we have seen is that the agent thread actually can replace the PR. And if you think about it, PRs are also so dumb because they're dead. You cannot try the app. You cannot edit it. You cannot ask an agent questions in line unless you got like all those dumb code review bots that are like GitHub comments clearly like, you know, horseless carriage style. So for us, whenever we want someone—it's not that we never want feedback from someone else on the team—but we can just invite them into the agent thread.
So, you know, I guess this example first, it's, you know, a change where it might have had an issue and you can, you know, say let's ship it, let's monitor the rollout. Turns out in this case, this is a real thread, there were some bugs, it hit some issues in the logs and then we reverted it. And I'll show another one where we bring some more folks in.
We've also instrumented this entire flow from, you know, people using our product, hitting an issue, reporting a bug. We can then investigate it. We get this diagnostics bundle with everything. Has anyone here reported a bug in AMP? Got an email. All right. So, you know, this is what we see on our side. And if we click, you know, 'I'm feeling pucky' or 'I'm feeling ultra pucky'—that's our little mascot guy—then it's going to go start an AMP thread with all the diagnostics, all your, you know, thread data for that thread only because that's what gives us access to see it all, the logs, you know, everything. And then when we, you know, here's what it looks like. It's going to go and figure out the issue and then we can ship the fix and it's pretty delightful to users to have something fixed in, you know, sometimes 15 minutes or less.
We'll send them an email because we've also let the agent send emails to our users. We do always have mandatory human review and confirmation of that email text. So, you know, we'll send an email: 'Hi, Alex. Thanks for the detailed bug report,' and we'll give them free credits and close the loop there. We don't have a support team. I think, you know, obviously we don't want to go and add that and we've really molded the agent to work with this. So, we all use AMP obviously. So, we can invest a little bit more into AMP. We can make it so it can do all these things like check our logs, like our bug reporting support system. It's all integrated in AMP and it can email and it can grant free credits and all of this stuff. If there's an incident, it knows exactly how to go look at our Grafana and see is like what is broken.
There is an advantage that we get because we are all using the same agent. And over at OpenAI they're all using Codex, at Anthropic they're all using Claude Code. But sometimes it feels like a lot of other companies are like they have like seven different agents and they're constantly switching. And sometimes it's because of token limits or spend limits, but you get so much more out of the agent if you just use one. And I know this like in theory the most self-serving thing ever. You don't have to pick AMP. I'm just telling you any agent is better than like seven agents because then you get to build all these automations, really build your company around the agent.
And one of the things that is particularly cool is multiplayer when everyone is in the same agent. I, you know, I think a lot of these other ones don't have it yet. They're probably going to add it, but in AMP, we've got the ability to bring other people in and to, you know, say, 'Hey, here's an issue. Someone else, can you look in this?' And this is what used to happen in Slack which is disconnected or it used to happen in the PR which just, you know, sucks. You can have a conversation here. Every person can also ask the agent for help and then you can make the decision here. And if you're in like seven different agents or even worse if your agent is something that runs on your own machine, then you cannot bring other people in. So if you realize all of a sudden you need help from someone else on your team, you're kind of stuck.
I think that's the future. And then going one level higher, what does the company of the future look like? We, because we're 21 people, because, you know, we're like profitable and all that, we've had the luxury to like do crazier stuff here. You know, we're 21 people. This message came out last night. I did not time it for this, but this is exactly how we think that the wave that is coming of AI and agents is going to completely change the way that companies are structured and completely change how they work and change what the role of the human is. And if you don't follow that through to its logical conclusion, then you're going to be whipped by someone and by some other company that is. So with 21 co-founders, we don't have to worry about token spend. We don't have to worry about, oh, do we trust this person? Every single person, you know, joined, we've worked with them for at least a year, in some cases 15 years. Feels more like a, you know, tightly held family business or something or like a hedge fund or something like that.
In tech, what sucks is every company is either like growing by 50% or cutting. And I used to, you know, as CEO, I would think like psychological safety, what the hell is that? It sounds like the dumbest thing ever. I think like what it means is everyone on the team now they can like try, they can fail, they can do experiments, they can disagree with anyone and it is the most open thing ever. And if you're a company that got really big in 2022, I think you're really going to struggle with that. If you're a company that, you know, maybe started in the last year and you're going and hiring a bunch of people that are for this one role and that all joined at different times, that don't have a big amount of skin in the game, that don't own a big chunk of the company, they're not going to feel that comfort. And I think you're going to be whipped by another company that looks more like this where every single person cares. If you recall back to that slide earlier, you know, basically everyone is working on the weekends. Everyone is, you know, working really hard here. And I think, you know, I welcome competition from our competitors that have 30, 40, 50 times the staff and same amount of revenue.
We also do think that software is dead. We haven't quite seen the bodies yet. There's a lot of workflow tools that obviously seem wrong, but we are building all kinds of software. I mean, this right here is software that we built custom for making slides and it's got all this. There we go. Get back there. Here we go. You can see all the slides here.
And making new software for yourself is so easy. It's cool to like, you know, play with random stuff on the side, but it's even cooler when you can build it into how you work constantly. So, that's a hackathon idea. You can even remix these slides if you like this app. Just go to amp.store and you'll see that it's kind of like a little experiment that we got going.
And I think that software, you know, what does it mean for software to be dead? It's less about the actual packages. It's more about all the smart people that have been locked up in software, all these IQ points for the last 20 years. And I don't think that that's going to be the case anymore. But all these smart people, they're going to be figuring stuff out in the real economy, in manufacturing companies, in finance, in companies that build real stuff. And it was a historical aberration that people could be building software for other software companies. That time is dead and anyone clinging on to it is probably not going to be the agent-pill people that you want to be working with. So I'm actually really excited about this transition.
Let's check on— Yeah, clap. You want— someone want to clap? Thank you.
All right. So this change has pushed to main. Pretty soon you'll be able to go and, you know, just type in any thread in AMP and if it has 'Frontiers 2026', then you'll see a nice thing with some free credits. So, I wish it had gone a little bit faster, but I said give the agent a long leash. So, we like to have it do a little bit more verification of its work. Do we have time for questions?
