Most of Your Docs Readers Are Now AI Agents - Hahnbee Lee, Mintlify & Lincoln Murr, Coinbase

AI Engineer · 18 min · 193 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Hi.
00:13Nice to meet you all.
00:15Um, appreciate you guys all coming.
00:16Um, hope everyone is having a good start to their day.
00:19Um, nice to meet you.
00:20I'm Hamby.
00:21I'm a co-founder and CTO at Mintlify.
00:24Uh, Mintlify is a series B company based in San Francisco and we power developer documentation for companies like Coinbase.
00:33Um, would you like to give an intro?
00:35Yeah, sure.
00:36Everyone, my name is Lincoln.
00:37I'm the head of AI product at Coinbase, uh, which you might know as a crypto company, but is widely expanding into all sorts of assets and uh, payments as well and there's also an agentic angle, which we'll be talking about here shortly.
00:49Yeah.
00:50Um, do you want to go and ask the first question?
00:53Yeah, sure.
00:54So, the format was just kind of kind of fireside chat back and forth, talk a little bit about how how we view the world today, how our respective companies are building for this agentic future and where we kind of think things will will go.
01:03So, Hamby, can you kick us off with uh, with a question here?
01:09Sorry, I'm just Oh, yeah, yeah, yeah.
01:12This one.
01:12So, yeah, how how do we think about agents becoming the the primary actor on the internet and the primary thing that people are using as as part of the uh, their internet experience?
01:21I can go tell an agent to go set up Coinbase for agents.
01:24It's going to find our docs and do it all entirely for me.
01:27How does this change what's important when you're thinking about how you're building docs?
01:30Yeah, of course.
01:32Um, so, Mintlify was founded in 2022, so pre-chat chat GPT, uh, GPT-3 specifically, I would say.
01:41Um, and I would say that previously there was like much more of an emphasis on design and front end, which I personally love design and we are definitely like a design-led company still.
01:52So, it's kind of still it kind of makes me sad.
01:54I'm not going to lie, the fact that like the agents do not see your website and they do not care what it looks like or what the branding of it.
02:01And so, what ends up mattering a lot more is like things like page size.
02:06Like, is it taking up a bunch of like the context window?
02:09Uh like the the content being accurate because like an agent can like set like read your docs and set it up, and then if it runs into like a little friction point, it's just going to stop.
02:19It's just like it's not going to like debug it.
02:22It's It might just like It might try and try to like brute force it, but then it can also just as easily churn to another product.
02:29And so, I do think like, you know, like coding agents have made the world like move so much faster, and it's important that like I like engineering or EPD specifically
02:41is like seeing so many like uh like productivity gains, and uh like the rest of the org kind of needs to catch up, you know?
02:49And I I think the only way in order for us to all move at the speed that like the productivity gains that engineering is seeing is to kind of use AI to kind of keep up.
02:59And that means like, of course, it's a lot more subjective like whether the docs are good or like, you know, writing AI writing I I like I personally like I'm still figuring that out.
03:10Um but I do think like setting up automations in order to ensure that like uh like things like do not go stale is like a very very important.
03:18Um and I'm curious for yourself.
03:20So, now that like uh agents are making transactions on the internet, how does that change how Coinbase thinks about uh like the like like just transactions in general?
03:32[laughter]
03:33Yeah, the the the intersection of AI agents and and stablecoins as a as a payment rail have sort of happened at the same time.
03:40And traditionally, you see two inflection points come together in in previous eras of technology, like for example, mobile phones at the same time that you have 4G LTE.
03:48And this opens up a lot of new opportunities, but it also opens up an opportunity for us to be disintermediated or disrupted by a new nascent startup or company.
03:57So, there's there's really two ways that we think about it.
04:00One is what are these innovative products that we can build and use uh and and market specifically toward agents and build for agents so that they have this experience that is otherwise completely not possible in the previous era of the internet.
04:13But then also too, what are the core parts of Coinbase's business that we need to make accessible to agents such that people continue using Coinbase as for our bread and butter use case of of trading and money management and whatnot.
04:24And the way that I view those two things is around how we make our entire business accessible via APIs.
04:31So, how can we take all of the functionality that's available to humans and translate that into a world where agents can interact with it.
04:39And though that's possible with browser use, it's not exactly the most ideal situation.
04:43There's a lot of authentication and and security concerns, especially when you're dealing with finances.
04:47You don't want to give your agent sort of unfettered access to your account.
04:50You want the right guardrails, limitations, all of that in place.
04:53So, to to to build this defensive hedge for our company and allow us to survive in the AI era, we are taking this it's sort of an analog to
05:04web shifting to mobile, but now mobile shifting to agents and making everything that was accessible for mobile users inside their their phone back a decade ago when people were saying, "I would never bank on my phone.
05:14That sounds crazy."
05:15Now everyone does it more than they do even on their on their web browser.
05:19Doing that same thing, but now making it available to agents adding in all of the safety nuances that you would need such that today you may say, "Why would I ever give an AI access to my account to manage money?"
05:28And 10 years from now, you won't even be looking at your account at all, just be happening through your through your agent.
05:32I'm curious, is there like specific advantages of crypto being the solution for this?
05:38I think on the payment side, absolutely.
05:40So, a lot of what we've seen with early traction around X402, which is our internet native payment standard that we have now externalized into its own separate foundation.
05:49That's that's co-run and governed by Cloudflare, Stripe, Amex, Mastercard, Visa, and so on.
05:54The the crypto angle, I think is primarily around the ability for digital dollars to be sent uh cheap, fast, permissionlessly anywhere in the world.
06:02So, anyone who's running a business and wants their AI agent to accept a transaction can do so.
06:07And though we've seen this a lot with existing sort of card-based or stripe link style solutions, one of the areas where I think crypto has a particular role to play is around micro transactions
06:18and the ability for an agent to spend 1/1000 of a cent, even 1/1000000 of a cent, to access an API or a service or even a paywalled website very, very seamlessly and and trivially.
06:30And I'm curious, is that future already here?
06:32Like are we already seeing agents making transactions via crypto?
06:35Like
06:37Yeah, absolutely.
06:37We're we're seeing a pretty exponential increase in the amount of transactions that are being made.
06:43In the past year, we've seen around 100 million agentic transactions actually, all happening on chain, all publicly available for anyone to audit.
06:51Uh all of those are are primarily for this use case of of interacting with APIs.
06:55So in this previous version of the world, you would have to go to whatever website you wanted to interact with the service, you'd have to fund that account, you'd have to get an API key, and then hand that to your agent.
07:05Only then you can really start to interact with things.
07:07But now we're seeing a world where if you just give an agent a wallet or some sort of financial account, it can go off and pay in a one-off basis for different APIs.
07:14And that's been a really resonant use case with with this core audience of of early adopters of agentic payments.
07:21So you're talking about like basically people connecting to your wallet via API or like the agent purchasing APIs or
07:30Yeah, it's like the agent purchasing APIs.
07:31So
07:32Oh, okay.
07:32The the analogy I like to use is it'd be like if every the the current way the internet works is every time you uh it'd be like if every time you went to a grocery store, you had to walk in, you had to buy a gift card for that specific grocery store, you had to sign up for
07:45their loyalty account, and only then you can start paying for stuff.
07:48And that's how it feels whenever you go to a website, you have to get API keys, you have to fund that account specifically, and only then your agent can start interacting.
07:54But as we give these agents more financial independence and autonomy, they're able to do things without having to go through this arduous human blocking process.
08:02And in the same way that I can only do so much on the internet without having access to money, agents are very much the same way.
08:07Giving them access to money is how you increase their autonomy and ultimately get to the point where they can go do things if you just tell your cloud agent, I want to go generate an image, it should be able to do that even if it doesn't have a connection to that today.
08:17It should just pay for that service, pay the 25 cents or whatever it might be.
08:21So, Coinbase is like doubling down on APIs in general, would you say, to make it more agent friendly?
08:27Yeah, absolutely.
08:28I think APIs are the name of the game.
08:29You can wrap that in an MCP and that's all fine and dandy if that changes at some point in the future.
08:33We'll see, but uh it's it's really the the the APIs and that underlying uh access to your service in a way that gives makes it so much more accessible to people in like a safe, friendly, intuitive manner.
08:46I was going to ask if you have any takes on like MCP versus CLI versus API.
08:51Yeah, I just uh I think API's always going to be that fundamental layer underneath it all and then the MCP is nice for adding additional context and nuance that's required within like a natural language style setting and and CLI is is broadly
09:03the same with the additional requirement of having like a terminal associated with it, but between all of them, I think it's kind of like a you have to have all of them and it's how easy AI assisted development has gotten.
09:12It kind of feels like you're just fundamentally wrapping them in in whatever way, shape, or form that the agent can can use.
09:17And beyond that, there's, you know, pros and cons to each of them individually, but I think we're we're far from the end game of how these tools evolve and and where they start going.
09:24Yeah, that leads to my next question.
09:25Like, where do you think the world is going to be headed towards?
09:29Mhm.
09:30Mhm.
09:30[clears throat]
09:30I On the agentic payment side particularly, my my sort of core fundamental belief is that granting agents more financial independence by giving them access to money and and authorizing them to spend that is how we create increasing levels of autonomy and long-running agents that just lead ultimately to better outcomes.
09:48If I can get my agent uh the ability to go pay for information on paywalled sites that otherwise wouldn't previously be available, that would allow it to run overnight and do even deeper research settings that I could do if I just gave it access to the public open internet.
10:02Similarly, if I allow my agent to in a in a future version of the world, say I want my agent to go set up a company for me and form an LLC.
10:10Maybe it can't do that entirely on its own, but maybe it can go hire a lawyer off of a you know a Fiverr equivalent or or rent a human or any one of these services and and an accountant to handle this work.
10:19It there's a really interesting concept from the crypto space.
10:21It's called a DAO like a DAO.
10:23It's a decentralized autonomous organization.
10:26And this has been around for for several years now, but it's really manifested itself as basically just another form of bureaucracy where there's a handful of people trying to make decisions and it just it just always kind of crumbled.
10:37It never really achieved that whole vision.
10:39And the original vision was automation at the core and humans at the edges and you run an organization that way.
10:44And I think agents with access to money are a very clear example of how we're going to see that vision finally come to life.
10:49I genuinely ask these questions with curiosity cuz I personally do not know crypto very well and it's very interesting to see how like I guess like a lot of these theses that were like kind of built previously
11:02are kind of coming into fruition with AI in play.
11:05Yeah, it's been really fun to see.
11:07It's very validated in a lot of senses where it's like you simply can't just give an AI access to a bank account.
11:12A crypto wallet is a nice way to provably give it access and and ownership over over money, but I would love to hear your your perspective and vision on how the world evolves as well as as agents becoming the primary
11:22adopters and ingesters of of documentation and how that changes your strategy at Menthafy.
11:27Yeah, I love how much of an emphasis you had on APIs cuz that's also my take.
11:32Like it's it feels like a lot of software companies are kind of becoming more developer-centric.
11:37And you know, we do envision a world where everybody is using like this technology, but right now like the earliest adopters are developers and also the agents actually speak
11:46like the language of developers because like it's all in code, right?
11:51And like they they're really good at bash.
11:53They're really good at CLI commands.
11:55And so definitely think at least like for now like doing that and all like software companies should be really investing in APIs and uh these things that like seem a little bit traditional honestly.
12:09Um and also I I think like what you said also reinforces the fact that the like front end is not mattering as much either.
12:16Like make make everything avail- accessible to agents by like by creating APIs essentially.
12:24Um one interesting statistic that I want to share is that at the beginning of the year across all Netlify sites, we had around like 10% of like the viewership
12:34was agents and now uh we're closer to like 80%.
12:38Um and it's kind of funny we have like this internal tracker at our company where we uh we just track like how what what the percentage is and we're like we internally like take bets on like where we think it's going to go and it's been growing faster than we even think.
12:53And it's not like human viewership is going down, but I think that like the agent viewership is just going so much higher.
13:00Um I always like to like it's it's weird because like agents act kind of similarly to humans and uh and different in in many different ways, but like I like kind of using like the agent terms on humans in the sense that like we as humans have very small context windows
13:16in comparison to the agents like and they have much more like you know, time to just like like ingest so much more content and be able to grok that well and I think that's like really the the crux of like the big surge that we're all just like seeing right now with like how agents are impacting businesses
13:33and like how we work.
13:35Um it's it's it's insane like the like the the growth like over the past years is insane.
13:41Uh like uh open claw like was literally like this year and that's that also kind of feels like forever ago.
13:47That that's fascinating.
13:48I want to drill in on something you said about uh companies investing in like the traditional things and also what you made previously around humans writing documentation as opposed to AI.
13:57I'm really curious to learn a little bit more about your perspective on that and if we'll always sort of need this human touch to to write documentation or or or provide a nuance that then an agent can go ingest
14:08or if there's a world where this starts to become a a more automated process for teams and how teams are interacting with this uh this uh pattern today.
14:14Yeah, I spend a lot of time thinking about like human writing versus like AI-generated writing and where my takes fall in that.
14:23And I do think that for creative prose or narrative writing, uh it's very important to have that human touch.
14:30And I don't know how terminally online this room is, but I am quite terminally online.
14:35And there is so much conversation about uh you know, like uh software engineers like using uh coding for writing all their code, but then religiously like writing like uh prose by hand.
14:49Like, you know, when you're like giving a document to somebody else within the team, like do you decide to AI-generate it or do you decide to write it yourself?
14:57There's starting to become like a class of documents and like division happening now.
15:02And so, I definitely think like um it's I like I always kind of like uh teeter around like, you know, going towards tradition versus like going towards like a new paradigm of like everything is changing and we need to change the way that we think.
15:16And I think that it's we're all figuring it out as we go.
15:20I I I'm not going to lie, I do think I had a point in time where I was like feeling a little AI psychosis, where I'm like, oh, like, I don't need to talk to anyone.
15:28Like, I'll just talk to my MCPs and I'll be able to get all the insights and everything.
15:32I definitely think, especially as a product person, I'm also curious to hear your thoughts on that.
15:36Like, it's still important to talk to people and figure out like what what what the real pain points and use cases are today uh in spite of all the innovations
15:44that are changing.
15:46Um but uh yeah, to to kind of go back to your question, um like I think for technical writing particularly, like there is like a process of gathering the insights
15:58and then like putting it in, in which I do think like the technical writing like portion of it can be automated.
16:06Like, I think it's very tedious to find all the related PRs, find all the related context, all everything that's happening within Slack, all the decisions being made, and then like put that into something succinct, and then figure out like where
16:20uh and like amidst your corpus of like content like you need to like inject it in.
16:25I think like all of that can be automated.
16:27Interesting.
16:28Yeah, I I I've noticed from the product perspective to your point that it's it's very difficult to find any sort of insights or takeaways from just interacting with an agent.
16:35And if anything it just hallucinates and make you prioritize the wrong wrong things.
16:38Um if my reckoning of time is right we probably have time for one more question.
16:42So I'd be I'd be very curious to hear from you what is the most common fit fitfall you see teams who are building for AI and building documentation for AI
16:50uh fall into and and and how that sort of impacts what they're attempting to drive the AI to do and what you would recommend to to make this whole process much more seamless.
16:59Mm yeah.
17:01Okay, this is going to sound like a a my own plug, but I do think that uh like it's the whole like build versus buy.
17:10I think is like the question that everyone is constantly asking for software products.
17:15And I do think that you know, when you're deciding what to do whether it's documentation or any other SAS that you have within your company to like improve your productivity
17:24or like not your main thing, I do think that there is like it's you're either deciding to maintain something yourself or uh like offload that to another team that's literally spending like every waking hour thinking about it.
17:38Um and so I do think like things are getting more competitive with AI, things are moving faster with AI and I think like every company should be focused on whatever their number one value proposition is and it's so easy to kind of forget
17:51about uh the maintenance costs of things, you know?
17:54I think that's like what we're kind of seeing a lot in enterprise sales as well where it's like "Hey guys, we built this during a hackathon and nobody wants to maintain it."
18:02Like what like like if we were in the business of doing what you guys are doing, we would be building your business, you know?
18:08Um and so I think that like uh really remembering and I think this has always been the case.
18:13I think just AI like accentuates this a lot more.
18:16It's like "Hey, we could build it, but like do we want to maintain it?"
18:20Um and I think that uh this is a question that we also ask ourselves when building products.
18:25Cuz we could be building everything.
18:26Yeah, that that makes a lot of sense.
18:27And especially with how fast-moving the agent space is and and and the way and manner in which they crawl sites and whatnot.
18:32It feels like there's a million different standards and whatnot.
18:33So, having just like a constant view on what the state of the art is and having that handled without having to do the maintenance.
18:39I've seen that firsthand with hackathon projects and whatnot.
18:41So, yeah, couldn't agree more.
18:43Awesome.
18:43Thanks so much for your time, Lincoln.
18:45Yeah, absolutely.
18:45Appreciate everyone uh sitting through this.
18:46Thank you.