Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
x402 Roundtable ft. SKALE and Trails
Recorded: Feb. 26, 2026 Duration: 1:06:38Space Recording
Full Transcription
Music Thank you. The Oh Thank you. Welcome, welcome, welcome.
We will be starting momentarily.
We're just having some technical difficulties on the back end.
No worries though.
There he is.
Let me invite him up here
hey hey James how you doing
hey how's it going
doing good glad you were able to sort out Hey, hey, James. How you doing? Hey, how's it going, Swearer?
Doing good, doing good.
Glad you were able to sort out the Twitter bugs.
I don't know what it is.
It's always either it works on the computer or works on the phone,
but not both of them at the same time.
So that's always the case.
Sounds about right.
What can you do, though, right?
It's just the way it goes. Yeah, maybe send up an agent and that can get a fix for me seriously we actually had our first
agent join last week we'll see we'll see if he joins this week the uh the dexter agent popped in
so who who knows at this point we could all just be agents and i don't think anyone will be able
to tell the difference so can they talk nowadays how's that going i don't think anyone will be able to tell the difference. Can they talk nowadays? How's that going?
I don't think anyone's gotten one to talk in a Twitter space yet,
although I'm sure it's possible.
I think once that happens, then it's game over.
Okay, I think that's a weekend project in the future.
I'm going to see with real-time voice and see if we can make something happen.
Shoot me a DM. I'm happy to play with real-time voice and see if we can make something happen Shoot me a DM, I'm happy
to play around with that for you
I've been wanting to try that new model from the video
apparently it does
real-time
listening and
speech
Yeah, I saw
Eleven Labs is also
really good
and if you're familiar with Mistral AI, they have a completely open source one,
and it speaks in real time, which is really, really nice.
Incredible. Well, we've got some toys to play around with, that's for sure.
Anyways, welcome, welcome, welcome, everybody to another X402 roundtable.
My name's Sawyer. I run Scales Developer Success Team.
And I'm joined here today by James, who I actually don't know what your...
I don't know if you got a new title or not. I'll let you introduce yourself. He works and built Trails, which recently got acquired as part of the Sequence acquisition with Polygon, who is also joining us here in the space, which is really exciting.
So, yeah, James, please introduce yourself and, you know, tell the world what's changed and going on on your side of the world.
I mean, as you can imagine, a lot has changed, but all for the better. So yeah, Polygon acquired Sequence, which, you know, was a, Trails was a part of that. I can get more into what Trails is exactly for the people who are regular listeners, but in short, it means, you know, more resources, more abilities to like build out and scale out, especially like on the agentic side.
But this has been a really great opportunity for us.
Still the head of DevRel here.
So that's been a smooth transition over to the head of DevRel Polygon.
So yeah, really just connecting with builders and, you know, a core focus of ours is working
on X402, working on Agendic payments in general.
So happy to talk more about that and the stuff we're doing there.
Absolutely love it. Absolutely love it.
Yeah, I mean, X402 is an interesting beast.
And I know, you know, we'll come back to X402 in a second,
but I know on the, you know, the trail side,
it's something you guys have been working on.
I want to say what at this point?
Eight months? Six months? Does that sound right?
Yeah, it's been a while in the making.
So it's very interesting because it's been years in the making
in the sense of it's using all the infrastructure we built
over many, many years and thousands of users
and millions of wallets using the stack now.
But I think ultimately it boiled down to like,
okay, what is a really strong use case
that resonates with the market?
And ultimately that's kind of what Trails is a result of.
And that resonates with a lot of other people as well.
Yeah, no, fair enough, fair enough.
For those of you that don't know Trails,
went live into production a couple of weeks ago now. Really incredible intent platform lets you basically do cross-chain anything. So payments, bridging, actions, export to payments. It's pretty incredible.
excited that we were able to get it to work with scale so you can actually move assets
from any other chain to scale with with trails which is really awesome. Big, big thank you
to James for helping make that happen since we were able to work and do some pretty cool
stuff with trails and using our native scale bridge. But if nothing else to add there James, I'd love to jump over into, you know,
X-402.
Yeah, let's do it.
I mean, I think, you know, trails is really relevant to X-402 and, you know, um, I would
broaden this out to just agentic payments.
Um, I definitely see like X-402's a subset, a very strong subset of that.
And, you know, we're doing a lot of work on adjuncting payments because ultimately we want more volume across the board.
I think we want a bigger slice of the pie and to bring more value on chain.
And so that's what it's all about.
And we see Trails playing a big role in that, X4R2 playing a big role in that, X402 playing a big role in that. And so we're devising, creating all the use cases around
what are the combinations of these that really resonate
and provide value for either people that are running agents
or developers that are building kind of like agentic platforms.
Yeah, and that makes a lot of sense.
You know, for everyone kind of jumping in here as we go,
X402 is an internet native payments protocol.
And it is, I think, I want to say officially like a year since Eric wrote the first, like the first draft of the spec.
I want to say it went live into production in like in March, I want to say.
Or went, yeah, went live with V1 in March,
or give or take, yeah, something like that.
I don't know, dates are off.
But anyways, the part that's relevant
is it's been around for about a year now.
And we've gone through, I'm going to call it, three phases.
Phase one was the early builders.
And, you know, that was kind of when we had people trying to figure out what to do with it and how to make it work.
Phase two, well, maybe four phases.
Phase two was the meme coin craze that happened in, I want to say, September.
And someone dropped a meme coin using X402 and it kind of blew up and we saw a bunch of, you know, kind of irrelevant volume
for a little bit. Yeah, I know, right? The irrelevant volume that everyone now looks back
at the chart and says, well, why is it not doing that? And the answer is, well, people were just
buying me the coins. But anyways, then we went through a little bit of a lull, which we can call
phase three. And I would say the last like three weeks, we've kind of been back into this kind of new phase of Agenda Commerce. And so the thing that it was originally kind of,
you know, built and designed for is actually starting to come to fruition. A lot of it,
I think, is thanks to OpenClaw, which has obviously taken the tech world by storm.
If you leave our little tech bubble, most people in the world have no idea what that is,
and that's okay. But at least in our corner of the world, it seems like OpenClaw has really
opened the eyes for a lot of people on how do we actually make our agents pay each other,
right? How do they do commerce together? So that's kind of a little bit of background.
We're going to go through just a bunch of random topics
that I think are interesting and relevant. James, at any point, feel free to just run with it,
throw out ideas, et cetera. This is supposed to be just kind of talking about all the different
things happening in the space. And of course, if anyone wants to raise their hand, come up and
share what they built recently, you're more than welcome to. But yeah, that's my little recap on
X402. And what I think is driving,
one of the key factors that I think is driving the recent interest in actual true agentic commerce.
James, thoughts?
Yeah, so I got a lot of thoughts.
I've been working a lot with agentic builders.
We've been building out our own,
you know, agentic kind of like platform as well.
I think, you know, everyone's launching a wallet nowadays and we're doing a lot of other stuff, you know, using our infrastructure to do that.
So we've got a really great launch coming up and yeah, we can share more about that later.
But I think in general, like the drive with X402 and what I see a lot of use cases is what makes me excited is people actually
making tangible payments. And you can see, you know, week by week and month by month,
the volume is slowly tricking up. So like we, as Polygon, we noticed that, you know,
X402 is a very strategic initiative. So we actually created like a governance proposal
in order to redirect a portion of like the fees that were being burned on Polygon in order to subsidize gas for the relayers.
So if people here aren't aware, basically you can execute an X402 payment without having gas in your actual wallet.
And generally that's just encapsulated within the fee that the facilitator is taking.
just like encapsulated within the fee that the facilitator is taking.
And in the case of Polygon, like if you pay in USDC to one of our facilitators or really any other ones, doesn't matter.
We're sponsoring like the gas on that side.
And so what's really interesting and that we're seeing a lot of is agents don't have a switching cost like humans do.
don't have a switching cost like humans do.
So they will naturally gravitate towards what is the path
of lowest resistance in terms of fees, in terms of
usage, and in terms of the amount of
volume they can actually use if they're doing a swap or a transaction.
And so we see a lot of when you're matching fees or you're
reducing fees, the agents naturally gravitate to you.
So that's one thing that we're seeing a lot of is agents paying in next 402.
And I think another thing that I'm seeing a lot of personally, you know, as part of Polygon, of course, there's Polymarket.
And I think agents are in a very unique position here because you don't have, like, basically agents are still really bad at time series forecasting.
Like you see this in like a lot of quant firms.
They're still not using LLMs to their full extent for analyzing the market.
But LLMs are extremely good at reading expos, reading what's happening in the market and reading kind of like what people are talking about and then acting on this.
And like prediction markets are such a great venue in order to basically extract that information and then to get an advantage on a specific market or a trade.
So we see a lot of people deploying bots specifically in order to capitalize on this.
But we see a lot of people deploying bots specifically in order to capitalize on this.
And I think that's kind of a really interesting thing that's unique to crypto of people using
prediction markets in order to like self-fund and pay for the tokens of the LLM.
So there's some exciting stuff there.
Definitely.
I very much agree with you in regards to kind of where we're seeing a significant amount of the,
where a lot of the agents are being pointed.
Obviously trading in the crypto space
is always going to be probably our number one use case,
even though in the more broader sense of the world,
I would say trading agents are probably a small fraction
of what's actually being utilized.
To your point, they're just not that good by themselves.
If you feed them enough data and you're really good at managing context,
they're really good at how you put guardrails around them.
You can do some pretty incredible things,
but we've seen a lot of the trading examples be primarily pretty manual,
not too much autonomy.
The part though that I want to pull out there that I think is really cool
and that we're starting to, I think, see more of is this kind of, I'm going to take care of myself
mentality that we're giving to agents, right? So it's really easy to create an agent that runs in
the equivalent of a well loop and has some directions and you give it some API keys and it will theoretically
run forever, right? Because, well, you're funding it. But the part that gets really
cool and where I think X402 becomes really interesting is, well, I don't necessarily
want to let this agent run wild and spend all my money. So if I put some guardrails in
place and tell it, hey, you have to actually make money to buy your own compute,
now it actually has to use its earnings and then go find LLM providers available via the bazaar or some form of discovery layer
and pay them for that inference, right?
And so it becomes a lot more of a dynamic system, which has, I think, a dual benefit.
One, you can technically launch a lot more agents.
You don't have to worry about getting destroyed from an X perspective.
If the agent is good, it should, in theory, pay for itself and its operations.
But two, it also means that your agent can, over time, potentially be far more cost effective
and actually have a higher ROI for you.
Because maybe it can find some inference on sale or on some discount via the bazaar,
which should ideally be updating live. That's another topic we can chat about later.
But it's just really interesting how we're so early, but one of the key use cases here is quite literally X402 enables autonomous agents in a way where without it, it would just be so much harder because it wouldn't be truly autonomous.
You'd still have to sit there and ensure you pay your credit card bill.
Whereas in this case, you can fund an agent and well, it could succeed or it could run out of money and do the equivalent of dying.
I'd hope for the best.
I do love these.
I know like Anthropic does a lot of these experiments.
And I saw another paper that someone published where they basically just let loose 20 different agents from 20 different model providers.
And just had them basically compete against each other of like who could basically outsmart who or who could exploit who.
Essentially, they just deviate into degenerate behavior like very, very fast,
like trying to get the most money and trying to attract the most value,
which is a very interesting like right now, I think there's still a huge kind of like information gap.
I think there's still a huge kind of like information gap.
Like even I, you have to keep up on the space
so much every day because it's changing so fast.
And what I kind of notice is like,
you fall behind really quickly
if you're not keeping up to date on what are the latest skills,
what are the latest deployments that are happening,
what are the latest ways to like orchestrate
and do this scale.
And I think this is what's very interesting is like,
we still haven't, like, agentic commerce in X402
is still not verbose or distributed across a lot of players.
And so I wonder when we hit this threshold of like,
everyone has an agent or everyone has 10, 100 agents,
and they're trying to find these different like gaps in the market in order to pay for their compute,
you know, will this still be feasible?
Because I think there's like this gap where you can, it's not yet capitalized on by a large number of people.
Definitely. I very much so agree.
And I think a big limiter here is discovery.
I think there's an interesting question around, will discovery be free or not?
And how do you actually have good discovery?
I think a good example would be you've got the core V2 discovery layer, which is not
really being utilized by most facilitators.
And so therefore, there really is no generic discovery right now.
I would say majority of the X402 consumption
that I've seen has come from either like X402 scan
or from like native CLIs.
Like you saw Alchemy, right, launch their MCP today.
And so they're starting to embed these types of things
directly into their own tools.
But there really isn't a generic, like I'm to, you know, just as it comes, I'm
going to pay and consume.
We haven't seen a ton of that in, you know, from what I've seen.
And yeah.
It's so interesting.
This was actually one of, like, before I was talking about topics I want to talk about.
Discovery is an obvious gap to me that is frustrating because, like I said, I mean, I'm constantly
in the space and like you're on X, you're on LinkedIn, you're, you know, reading different,
like everyone that's pushing out different content and like what is the latest and greatest
or what are new skills to add. But it's so hard to keep up with where new things are deployed
and whether it's like something that's using X4R2 or something's something that's using X4R2
or something that you can buy using X4R2
or an integrator that's incorporated this
or a skill on Purcell's skills.sh file or in ClawHub
or whatever the case is, it's so hard to find.
And I was helping someone actually use the ReMotion skill.
I don't know if you've ever used this,
but it's great for creating technical videos
that are really pretty.
I'm a designer, and it's amazing for using this on an agent.
And I was talking through this, and it's like,
unless you have this knowledge that this skill exists
and where to find it,
it's impossible to actually incorporate this into your workflow.
So I think that's kind of missing is,
as I've onboarded a bunch of people into using agents
or deploying like their own, you know, claw, open claw,
you, they kind of like, they get started
and then like, okay, well now what?
Now what can I actually do?
What can I actually buy?
And I feel like agents are missing this autonomous layer
where it just immediately kicks off things that are valuable to them
and just finds skills or X4 or two endpoints
that are relevant to the human that's deploying the agent.
Definitely. I agree.
And like I said, I think what's really tricky right now is
majority of the resources
dead. Like a lot of the
X40 resources get spun up by people
and they just die.
And so there's actually not that many
that are legitimate.
Which is okay. I mean that's
kind of the way it goes. But I think
you've got teams like Zoth
apologies if I'm saying that wrong but, that's kind of the way it goes. But I think, you know, like you've got teams like Zoff or Zoff, apologies,
I'm saying that wrong.
But I think they've done a good job
of like helping validate
whether resources are alive.
But that doesn't help,
I agree with,
like what should an agent actually do?
And how does the agent start?
I think one of the,
I think there's a very interesting
bifurcation between like a human in the loop, a gender call, aka I'm using Claude code, versus I'm setting up a claw.
Just because the different things have different value on the actual call, right? Like if I give Claude code a bunch of skills,
it's pretty good at finding them depending on the model you use. But the agents.md is
far superior. Like the actual system prompt that you're driving into it is where you derive
the most value. And so I'm curious if we'll see some change in how these tools start to do almost the equivalent of like learn themselves, right?
Like obviously the OpenClaw bot, it's got the Sol.md and over time it can change itself.
Do we see, do you see a change in, like, I guess, let me throw this out as an idea.
I set up a claw or I set up some sort of agent.
I need some sort of like initialization command to say, here's what I want to do.
And then I need some sort of service that is curating all these services and things for me. And it's also, it's almost creating like the skill or the system prompt on the fly to give
my agent the best starting point to kick off.
And then it will continue to modify that over time.
But there is no zero to one starting point.
Right now, you like, like Remotion is a great example.
Remotion is really cool but actually like finding documentation and like how to use it and like get it into get it into
an agent was pretty painful um and then like prompts to actually know what to do was kind of
ugly um so can we make that a lot easier and does that then kick off discovery i don't know yeah
that's exactly um kind of along the lines
of what I was thinking is whether there's
like an aggregated discovery layer,
because everyone has their own like skills.sh
kind of repository now across a bunch
of different like platforms.
Or I much prefer similar to what you're talking,
someone posted a little bit like feedback
on how they're starting
up, in this case, not Cloud Code or OpenClaw, but Cloud Desktop.
And it was still confusing for a lot of non-technical people to actually get started and be like,
okay, what is a workflow for Cloud Cowork that adds value to my normal working day?
And a lot of people had really trouble, like, okay, what actually should I automate and
how powerful this is?
And it kind of did like, I don't know if you do this for like building PRDs, but you can
like do a ask question answer to Claude or whatever, and it'll like interrogate you of
the feature that you want.
And I love this way of building a product,
but I think it's also a really interesting way of like,
how do you build an agent that's custom to you
or that's useful to you and does the things you want to do
and it's specifically tailored for this?
Like, I don't know, maybe it's really, really good
at navigating Amazon, for instance,
or buying certain things from Stripe
or, you know, does whatever the case is.
I think eventually that this kind of custom formatting or this custom flow will get a
lot more refined.
And so you can actually have very tailored agents and it will automatically input in
and digest in an agent.md file that has all these different scopes and so knows in a streamlined way of what
to actually put into its context.
Definitely.
That makes a lot of sense.
I think maybe one of the tricky parts and maybe why this hasn't become so big yet is
I do think there's a really interesting kind of split around, I think, X402, where not everything
needs to be X402, obviously.
X402 is more of just a...
X402 shines the most when you don't yet know what you want to buy.
That's like the best, I think, best proof point yet is I don't yet know what I need,
but I kind of know where I want to go. And so I, yeah, I think like being able to figure out how
to seed an agent to know where to go, but not necessarily have to pre-seed it with every single
tool could be really cool. Especially because we are seeing a nice shift toward multi-tool systems, where it kind of does like dynamic discovery on the fly,
but it's still not necessarily driving an agent to figure out how to,
you know, how to maybe start with commerce out of the gate, which...
Yeah.
Oh, go ahead. Sorry.
No, I was just saying like a lot of times, it's very interesting,
because what you see on X of, like, people, like, posting of what their agent is doing,
and you try to replicate the exact same flow, and it takes a lot of manual work,
generally, to get, like, a full end-to-end from scratch,
unless your agent is already set up to accomplish some of this.
Definitely. Definitely.
I saw this with C claw when it was like blowing
up initially and people like rightfully so it's really cool and then like now half the time i try
to talk to it on telegram and it just keeps like disconnecting and i'm like yeah i i was uh i was
i was sitting down with ben miller over a zoom last week and we were trying to set up on an open
claw and honestly i, I was just struggling
and I'm a pretty technical person.
I was just struggling to get Telegram connected to it.
It was like the pairing thing
just like didn't want to work.
And I was just like,
oh, this is brutal.
This is miserable.
So I do think we're going to see,
which we already kind of are, right?
We've got KiloClaw.
We've got Minimax Agent.
We've got, who else just dropped one?
New Research just dropped an agent yesterday.
Airtable dropped one.
Manus dropped one a couple of weeks ago.
Who got acquired by Meta?
We are seeing a surge in hosted agents.
And I do think personally,
that will be the 99% long-term,
or at least the 80-20,
because I think it's just,
it's such a pain to set up the agent.
But more relevantly,
and this is where I think X402
is going to become far more valuable,
is I don't want to owe off
and do 100 services every time.
It's so annoying. Like, it's just so annoying. And, you know, setting up API keys and managing
different keys and billing limits and all this stuff is just, in my opinion,
an awful experience. And so I am pretty pumped to see a lot of these providers starting to offer it.
I am pretty pumped to see a lot of these providers starting to offer it.
I think the next step for X402 to really see hopefully a lot of growth is,
I think, getting the zero to one wallet guardrail on-ramp.
And then I think discovery.
I think we can kind of get like all four of those pieces really ironed out
within more of an agnostic provider.
I think we'll see X402 really just take off and go vertical.
Yeah, exactly.
I think we're like pretty close.
Like there's, you know, like I mentioned, like we're building a wallet and like we have like these guardrails already in.
Guardrail is already in.
Trails is really great for funding.
But I think for me, it's useful applications
is just the top of mind thing.
So I'm always looking for what's the next best thing.
I don't know.
I tried to run a human and these different applications
that are really interesting, which I think are quite cool.
But I haven't found like something
that really captivates me
and brings like that I want to direct an agent at
and have it run again and again.
And I think there's a lot of innovation now happening there.
So I think something will definitely come
where I can direct it to kind of make payments for me
or do something autonomously.
And we're seeing a lot of really nice use cases out of this.
Someone mentioned to me
that now it's essentially kind of like speed running
a seed round for a lot of these startups.
Instead of going through the whole brick and mortar roll of creating a pitch deck
and spending three months to raise,
you can just, you know,
go to a launchpad and launch a token
and connect it to an agent
and you're good to go
and then build an application from there.
Which is like,
how does this fit in like a more agentic world?
And how does every use case fit in this?
Definitely.
And I think like for me,
I've always kind of,
I think I've been saying this for a while.
So it probably is a bit more contrarian
than a lot of people
when it comes to the agentic commerce space.
But I do believe that you're going to consistently have
more of like your human to agent flows
and your agent to agent flows.
And I think the reality for
a lot of humans is it's hard to say, it's hard to say outside of a business perspective,
what you would necessarily want an agent to do for you. Um, because a lot of stuff is maybe like,
well, I don't, maybe I don't trust the agent to do it for me or like, I'm fine doing it. Um,
you know, me personally, like, do I love going to the grocery
store? No. But do I need an agent to constantly scan my refrigerator and, you know, figure out
based on my finances, how much I can spend, build my cart, order it and have it delivered my door?
Me, not so much. It's just, you know, not something I'm personally interested in.
But the part that I think is really cool is, to your point, as you think about
launching a project or launching a business or creating an idea, experimenting, being able to
spin up agents that can essentially go to work for you is really exciting. And where I think the
agentic commerce really becomes powerful is as we make the first, we're kind of in like this first iteration
of autonomous agents, right?
And I think as we continue to go forward
and we see these agents become more and more powerful,
I think you'll get into a hierarchy of agents
where you have agents spinning up more agents,
spinning up more agents.
I think that's where the agentic commerce
will really show itself.
When you're not managing that third or fourth or fifth or sixth agent down,
they're given a delegated wallet and it needs sign-off via multisig
from the person that created it or maybe the account manager.
And if they need inference or they need data or they need storage or compute,
well, they go and buy it according to the guard
rules that the other agent put in place for them.
That's kind of where I see things really going.
Yeah, I like the idea of the agent becoming more and more personalized to you, but also
more and more abstracted from specifically what I want.
Like, I think now this is an experiment.
Like, I would have my claw go to Amazon and go through the whole purchase flow and order something to my doorstep, which is, like, a fun experiment, and it's really cool.
But there's a lot of, like, edge cases there that you need to handle until it works very
smoothly.
But I would love to direct the agent, like, oh, I'm going on a trip to so-and-so.
And what should I pack?
And then it's just like, what should I buy in order for this?
And it just grabs everything, orders it to my doorstep, and handles a lot of the lower-level thinking.
Personally, I would love this. Maybe, I guess you don't love going to the grocery store, or you do.
I guess you don't love going to the grocery store or you do.
I, you know, I, I think it comes down to like,
I think Amazon's a good example of for a lot of things.
I don't necessarily, I don't need it to go the whole way for me,
but it would be nice if it went most of the way. Right. And I think,
I think like AP2 from Google has done a pretty good job of,
of kind of building that core flow where you can have different agents in charge of different things and they can wait for your sign off to actually execute the payment.
And that can all be rigged up with X402 or you could do, you know, I think their other framework, UPC or UCP.
There's a bunch of different ones. But I think that concept of like an agent can go do the work. And at the end of the day, you still have the approval to like yay or nay or redirect
it into something else, right?
So, hey, I need, you know, I need a new pair of shoes.
I like X, Y, Z or, you know, I need a new pair of shoes.
Go buy me a new pair of shoes.
It goes, looks through my Amazon, looks through my search history, says, okay, you bought
these four times.
You use way too many shoes, but I'm going to put these in your cardigan, and then you can approve or deny.
Even more shoes?
Yeah.
It's like, Amazon, you just ordered a fridge. Would you like another fridge?
Yeah, yeah.
But, you know, that's also back to, you mentioned something at the beginning about the speech, right?
And actually having an agent be able to maybe speak on here in real time.
That's also one of the things that I do think is going to become, I think, more and more
common.
I think once we hit a threshold of text to speech and speech to text and on the fly,
listen and response, being able to occur on commodity hardware,
aka on your phone, I do think we'll see an even bigger shift where, I use the Jarvis example.
If you haven't seen Iron Man, well, you should go watch Iron Man.
You'll see a great movie, but you're missing out regardless. But in this movie, he's got basically the equivalent of
what I think we consider to be AGI, right? A computer that lives everywhere. It can do pretty
much anything, build stuff, design stuff, speak back, process in real time, blah, blah, blah,
all the good stuff. And so I think that example we've seen a lot of people play with and experiment
with so far. I do think that for a lot of people is a good end goal where a lot of people play with and experiment with so far i do think that for a lot
of people is a good end goal where a ton of people already have you know like alexas or home pods or
i don't know what other listening devices in their houses heck even your phone could listen you is
listening to 24 7 just being able to be like hey hey xyz agent like i need this or you know you're
speaking and you're just like,
oh yeah, I forgot to do that.
And it's just like, does it?
I do think that's a big end goal
for the human to agent part.
Whether that requires agentic commerce or not,
I don't actually know.
But I do think from a more raw autonomous agent perspective,
I do think that's a lot of the end goal
for at least the consumer facing agent.
Yeah, once they get smart enough to
uh start charging us themselves then uh then we're in real trouble i think that's the that's
the threshold yeah yeah i mean we're we're in an unprecedented time right like if these things uh
if these agents start to start to do that to us we're we're in pretty big trouble so um
agents start to start to do that to us we're we're in pretty big trouble so um like it's such a
different even to six months ago i've delegated so much more of my work where like now it's a
function of how many tokens i can push out and i have like on one of my screens i have four
terminals more or less going all the time, just churning out content for different tasks.
And so in general, even the kind of work style is completely different.
And we've also leaned into this.
Everyone at Polygon has a $200 allocation for agendic stuff,
even outside of the core work.
And it's such a different way of yeah doing your business or like
working on different tasks where you're really like an orchestrator or you're like an administrator
of a bunch of like intelligently superhuman like employees that are just churning out code
constantly or churning out different tasks which is a very different way of working, actually, now.
For sure, for sure.
I mean, I don't know, I feel like
I rarely write code anymore.
It's primarily just like
if it really can't figure something out,
I'll go in and make the change, but
I feel like we're at the point where
Schemi 2.5,
BPT 5.3 codex,
Sonnet and Opus 4.6.
What else we got?
Minimax 2.5.
Minimax is really nice.
GLM-5.
I mean, we just have,
you have so many models to pick from
that are all like,
at the end of the day,
they're pretty much a better programmer
from a raw output perspective than almost everyone.
As long as you know how to speak to them,
you can get the output you want.
And I mean, sure, you may have to prompt them
a few extra times to get the polish you want,
but I mean, you don't need to write the code anymore.
You just have to know what you want your output to be.
So if you're an engineer and you know how to direct and guide,
you could be running a team of, I mean,
what feels like dozens of agents very effectively.
And I think we're starting to see that also bleed out
into non-coding areas, which is really cool.
And yeah, I mean, I also like, I do believe that is a lot of where, I think that's where
as a whole, I think agentic commerce will really thrive is domain experts are more valuable
than ever.
If you know how to do something really well as a human,
these models are trained on so much information.
If you know how to activate
the mixture of experts, experts,
better than other people,
there's a reason that an agent will,
instead of building the code itself,
choose to pay you three or $5
to build a small project for it.
Because you're going to be
more efficient be able to do it faster and they're almost like you right they can delegate out
hundreds of things at a time and so i do think that'll be a big uh i think it'll be a big shift
when we actually start to see that really come to fruition yeah that's what i i noticed that even
these models being incredibly intelligent they still don't have a lot of domain knowledge on very specific things and like very nuanced things.
I think like, for instance, like if I'm creating an invoice, I live in a place that like has a really long cycle in order to actually file your invoice with the government.
And it's super annoying every time I have to do it.
It would take like 30 minutes.
And I just automated this all with an agent.
But only because I've done this 100 times did I know all the steps that I need to take
and actually looking at the HTML that I need to parse out and passing in the keys that are necessary to authenticate.
And that's one really small tidbit, but I see so many instances of this when people are using
Claude and they try to use it in a way that they don't have super domain knowledge out,
and they just kind of take what it's creating as truth. like knowing how a system works or how process works is so important to actually getting a great output out of your model or your LLM.
Definitely. Definitely. I mean, I listened to a really cool podcast from some director of something at NVIDIA a few months back.
They got a lot of people who work there, but really, really smart guy.
Podcast was awesome.
I'd recommend it.
Go search NVIDIA mixture of experts
on your favorite podcast platform.
I'm sure you'll find it.
But they basically walk through
what a mixture of expert model is,
which the majority of the open source models
at this point are mixture of experts.
And basically what the mixture of experts is,
is instead of, if you have a huge model and what the mixture of experts is, is instead of,
if you have a huge model
and you have all of these parameters,
which are basically like the information
that it has available during its run,
the mixture of experts splits it into specific groups
where only groups that are needed are activated.
So for example,
if you're asking a coding question to,
you know, Minimax
2.5, it's only activating
some n number of tokens
compared to some
of the older models like the
LLAMA models, which activated
billions of parameters and took
forever to run. And so
it means that it may not necessarily have as
much information it's trained on,
but it makes it incredibly more efficient.
And so that ultimately, I think, is why we saw this surge in skills, right?
Skills are then able to help you bring in that domain knowledge.
But I think this is, again, back to just kind of this broader,
like, broader commerce perspective.
a broader commerce perspective,
people
sell their skills
for money.
That's what a business is.
And I think so far,
it's a lot of engineers
creating skills because
we're engineers and we don't really think about that a lot of times.
We're just like, let's go make cool stuff.
But I feel like we're going to get to a point
here where people stop open sourcing the skills
and they're just going to start making
specialized inference endpoints.
So like why would you open source?
Like Remotion is a great example.
The Remotion thing is so cool.
The Remotion library is open source.
Remotion could have not open sourced those skills.
They could have created a super awesome inference endpoint, and they could just charge a dollar
per shot using a cheap model capable of doing it, and 25 cents per iteration.
And they didn't have to open source that.
They did, but they could have made that design skill, again, more just like commerce directed, I guess would be the right language.
So I don't know if you agree or disagree there, but I feel like that at some point is going
to have to happen for this to be sustainable for non-LLM businesses.
Yeah, I was thinking through whether, you know, whether that's more valuable, like,
is it more valuable that this skill went viral for them?
And that like a whole class because like, to be honest, I've never used free motion
before.
And I probably never would have learned it if I never saw this skill.
It's all cool it was.
And I think if there was like a charge associated with it,
then maybe I never would have. So I think like the business model needs to be very thought through
of like, how do you get enough? Or like, do you have a lower tier model? I mean, similar to what
a lot of LLM providers do. Does this also like trickle down into skills that are payable via X402 payments,
where you can have a little taste of a pretty good model,
maybe for five requests or something,
and then it bumps up into,
okay, then if you want to go to the 4K model
or have a higher threshold model or the best skill possible,
then you have to pay for it,
which seems like a logical business model. But yeah,
I was thinking through kind of like that structure there. I had a question for you,
because I find this interesting, is do you think, I've been thinking through where people
will build like skills that allow agents to navigate the web natively.
And then there's the inverse of this, where you're a digital business,
and then you create either a native skill to your platform,
or you make it agentic-friendly, or you open up your API via an NCP,
or you just completely change your backend so you're completely api first and then make
your docs really good so agents can just parse it where do you see do you see more like agents
being designed to navigate the web like a human or the inverse where the people that design for
agents first will capture more of the market i think you'll see a little bit of both. I think
in the short term, we'll probably see more of the latter. No, the former. Sorry, you threw a lot
at me. We will see more of the agents navigating the web in the short term because it's what exists
and agents are already trained on web information, right? There's a lot more,
like this is kind of an irony for engineers, but most platforms historically have not had an API.
If you don't know what an API is, an API is basically like the backend that you can call
and do things, but you don't have necessarily an interface to interact with it. You can do it from
literally a command line and into like SDKs a lot of times, things like that. But there's not a lot
of APIs that are public. And I think for years, everyone has been chasing, trying to be as good
as some, like a group of very specific companies out there, aka like Stripe. They have one of
the best APIs ever, in my opinion. Come fight me on it.
But I think in the short term, it's a lot easier for agents to, yeah, to search the web and go do it because we've had these tools for a long time.
Playwright, Beautiful Soup, stuff like that, that actually literally are like for, you know, consuming the web as code.
Now, the part that's ironic is it's really expensive to do it.
And I don't mean just from a raw, like someone's paying for no reason, but HTML tags so far from the reports that have come out,
HTML tags consume an extra 80% to 90% of context in compared to just interacting directly with
APIs and markdown.
We actually saw, we saw Vercel and we saw Cloudflare.
Vercel actually did this first, I'm not mistaken.
But we saw them both two weeks ago or a week ago, create automated markdown conversion. So if you actually make a curl request to
a website hosted on Cloudflare or for sale or proxy through one of their services, their
CDNs, and you add the header accepts text markdown, it will actually return it back
just the markdown text. So they'll filter out all of the HTML tags for you. And like I said, it's like an estimated 80 to 90% reduction in context, right? Which means
you're literally saving money if you're paying for that LLM query over time. And so I think that
just goes to show that there's a cost benefit to doing it. The problem is the internet is not
markdown native. It's not native to that.
And so my opinion is long-term, you will see a bigger and bigger shift toward APIs,
toward CLIs, toward Markdown first or text first or some format that is highly efficient.
But I think that's more of an optimization thing that occurs for the hyperscalers. I think right now we're still
in the phase where it's kind of a, what can we do? And who's actually going to pay for this?
If everyone paid at cost for their LLM usage, I think we would see probably 80% of the industry
just drop off and stop using them.
They just wouldn't be able to afford them at the level they're actually using them.
Yeah. And so...
You can see the quad code, like using the API's 10x or 50x, whatever it is.
Yeah. I mean, I've got a bunch of different coding plans that I kind of
rotate between and the usage on some of them and
the output from some of them, it just, you know, they constantly vary because they're always
quantizing and doing different checkpoints and doing additional post training and things like
that. And so it's always a, there's always a shift, I think. But that's why I think right now
there's a bigger chase for what's actually the end goal for
people.
And again, I think it's kind of funny because agentic coding has obviously been arguably
the biggest use case outside of just raw chat.
If you go outside of the tech bubble, the number one AI application is still ChatGPT
from my manual polling of people.
I know a lot of people who barely know what AI is, but they know what ChatGPT is.
And they're like, I talked to this thing and it's incredible.
It's always so weird to talk to a developer who has only used ChatGPT or doesn't even
know what Claude is.
Oh yeah.
I always say, how does that even happen where you can't know about it?
We are, it's funny because there's obviously, and you guys know me, I don't do any market
talk.
This is not financial advice in any way.
I don't personally really believe in the whole like AI bubble.
Like I don't think these tools are ever going to disappear on us.
They're just too powerful and people love them too much. But I do think we're very much so in an early enough phase
where a lot of the AI companies will die and disappear
and a whole bunch of new ones will stand up
because we are trying to figure out what are they actually best for.
If they write all of our code, that's really great.
write all of our code, that's really great. But if it costs, you know, if it costs a business
$300 to, you know, build a piece of software, it costs 1000 bucks to build a framework with Opus.
Again, that's great. There's gonna be maintenance costs, there's human cost to actually ensuring
this thing is online and secure and all that good stuff. But then who's actually using it and why?
online and secure and all that good stuff.
But then who's actually using it and why?
And so I think again, I think it was the Y Combinator, they posted a chart last week,
or maybe it was earlier this week, about where LLMs are actually used, where AI is at.
And I think it was like tech is like 50% is like flooded with AI.
And as you go down, you get down down to legal and healthcare and stuff like that.
Their estimates are it's less than 1% of legal law firms and usage is being done by AI.
Yeah, I saw that chart.
That's wild to me that it's not significantly higher.
My assumption is that law would be one of the first things it would be.
You know, it's funny.
There's a startup called Harvey, which is also funny if you've seen the show Suits.
If you haven't, also you can watch it.
First few seasons are pretty good, but you'll get the joke and you go watch it.
But they're looking to push pretty hard and do that stuff.
But what I think is really...
I think one of the reasons that you probably haven't taken off is
at the end of the day, code and software
is theoretically a lower risk industry.
When you deploy a piece of code,
you've got a bunch of terms and conditions in your website
that basically says,
like, hey, I did my best.
And like, yeah, I did my best.
But like, when you actually go to a lawyer and you're like, hey, I need this thing.
Right?
Like, I need to be bulletproof.
Like, you're trusting a lawyer to do that for you.
a lawyer to do that for you. And so I think for the legal stuff, a lot of it is, like,
I'm sure a lot of the boilerplate is automated at this point. I've talked to a few lawyers,
it sounds like they've, they use it for some basics in a lot of cases, like just reproducing
like the same documents over and over. But a lot of that stuff I think is boilerplate to begin with.
I think the part where it gets really tricky is the nuance of the law.
Like what's actually legal versus not is people who have spent, you know, dozens of years
in their careers learning the ins and outs of the law, right?
And how do they actually work that to the benefit of their client, you know, in a legal
manner.
And I just don't think that's the type of thing that's necessarily easy to understand, because you would have to be trained on all of that data, which a lot of these LLMs probably are already.
But you have to be able to put two and two together, which is where I think people don't necessarily know how to activate those experts.
Yeah, that's an interesting point you brought up earlier, where, as mentioned, like you need a lot of domain expertise
in a specific sense. And I can imagine that there's, I actually like was curious and Googled
whether there's any like AI agent skills available, and I couldn't find any that were really good.
So it's very interesting, like I can imagine all the common software flows all have a bunch of
skills, they have a bunch of markdown files flows all have a bunch of skills.
They have a bunch of markdown files.
You have a lot of information out there, and so you can tap into it. But there's not a lot of lawyers that are spending time creating skill files for depositions or whatever the case is.
So I'm not a lawyer.
So that would be very interesting because there's not this domain expertise
that's probably been distributed across the web.
Definitely.
And like I said, that's where my opinion is.
I think that if you give away your domain expertise for free,
I won't say that you're not being smart.
I would just say that you're missing an opportunity. I think that there is a point where people are able to determine the cost versus the
risk of doing it themselves or paying for it.
And we see that all the time with SaaS.
The whole world is freaking out that SaaS is dead.
I promise you, SaaS is not dead.
People may be able to vibe code a CRM for their company.
You guys, I think I saw someone tweet
that you guys vibe coded a CRM for your company.
That's awesome.
I think if you're a super technical company
and you really want to maintain that,
sure, but the average company
probably doesn't want to worry about that
if it's not their competency.
And so back to the legal argument, there's a reason we pay,
like we sign up for enterprise agreements.
It's for legal coverage to ensure that they're taking the hit if something goes wrong.
We're not just storing a bunch of customer data
in a potentially insecure environment that was vibe-coded by a marketer.
Again, props to the marketer for building it.
Honestly, that's so incredible.
I can't believe I'm even saying that. Three months ago, that was not possible.
But it doesn't mean that everyone's just going to change their business to use only vibe-coded
tools. And so I think to the point of domain expertise, I don't think everyone is going to
try to build every skill and try to get LLMs to create their own skills themselves because there is a limit. At some point, if you want the best output, you have to
know the right words and the right information to trigger it. And I do think that will be
a potential catalyst for the growth of X402 when people start to sell that specialized inference.
Doesn't matter what model you're using, it's actually the information you can
inject into the system prompt that people are paying for.
Do you know of any
company
that's selling, for instance,
scale templates using XR2
or using any kind of protocol?
I'm curious if there's
any volume there.
I haven't seen it.
Honestly, I think part of the reason
is I think everyone... It is not an insult to everybody. I'm a bit of the same.
Everyone is a follower right now because we literally can't keep up because there's too much.
And so I think when it comes to raw LLM usage, my two cents of what's happening is
everyone is basically just like, the models are good enough, so I'm just going to kind of use free stuff.
And I'm not really going to worry about security.
Like people are just installing skills left and right,
no idea that they've got like scripts and stuff in there
that could just kind of crush you.
And I think you have real SaaS businesses.
Harvey's a great example, right?
Like they've got specialized inference now for legal.
And so I think those are the businesses that as a collective,
all of us together who are trying to build this agentic economy,
we need to go onboard those people to X4 or 2
and figure out how to get those people selling their specialized inference
because they're the ones that have it.
Speaking of, if there's any developers in the room,
someone go figure out a legal framework for X402.
I need a legal extension that allows me to do it
with terms and conditions on the fly.
Because if you're going to be selling legal inference,
you need a way to basically say like,
hey, use it at your own risk.
So someone think about an extension if you're a
developer. But yeah, that's my opinion is there's not too many people doing that right now. I think
we're seeing just a lot of like agentic trading, agentic finance. I think those domain models and
domain experts will come, especially as inference continues to get significantly cheaper month over
month over month. I'm hearing rumors that we're going to be getting...
I heard three new Chinese models
in the next week and a half
from various different tweets
that are usually right.
So every time that happens,
it forces the US model companies
to play catch-up
and launch as soon as they wanted.
They're going to launch the models
until the Chinese ice launches.
Yeah, I mean, they've always
got to be on top, right? So
I think like every iteration, we
see the compute get cheaper. As that happens,
it'll get easier to build the
domain experts. And
yeah, as that
continues to occur, I just think we're going to
continue to see the agentic commerce
grow to the point where
it may not even be an agent talking to an agent.
It could really just be like,
hey, I'm in chat GPT
and I have a question about
how to make a,
you know,
what do I need to create a C-corp
in the state of Delaware?
And there's a specialized skill
that like, for example,
Stripe created
and that's available
through their MCP inference and they, for example, Stripe created, and that's available through their
MCP inference and they charge you for it because it guides you step by step by step through
Or maybe they give that away for free because it drives you to Stripe Atlas.
Maybe someone else has one that you pay for and it uses Stripe Atlas as an example, but
you kind of see where I'm going with that.
Yeah, for sure.
That's definitely where the market goes. So I'm really excited to see.
And I feel like we're just the precipice and now it's building interesting applications.
And then what we've seen with agents, like we thought crypto moves fast, AI is moving
even faster. And I think it's just going to go exponential from here and take a lot of
crypto with it. So I think this is really beneficial for us.
I agree.
So let me, I know we're about at time,
actually we're over time.
This was a great conversation.
We're going to have you on again, for sure.
But I want to leave with one final question
before we wrap up,
because I think this is something that is very,
we'd say it could come off controversial. Do you believe that crypto
as an industry is going to get absorbed into the broader AI bucket? Or do you believe that crypto
will always have legs on its own? Oh, that's an interesting question.
I saw, just for context, I saw someone tweet about it the other day and I was like, ooh, that's
an interesting one.
So yeah.
Yeah, that is an interesting question.
I haven't thought about whether, you know, does crypto become a subset of AI in general?
My feeling right now is generally no.
And the reason, so basically I do think like AI
will accelerate crypto exponentially as I just said.
Because I feel like the strongest product market fit
for crypto now today is basically payments,
specifically like stablecoin payments,
and then DeFi as a secondary. And then you also have like, you know, other things like prediction
markets, so on and so forth. But in general, like payments are just such a strong product,
make it fit. And they just go so hand in hand with agents that I think this will just accelerate.
with agents that I think this will just accelerate.
Now, I don't think though that crypto will just kind of be absorbed or lost in AI
because I think they're quite distinct in a lot of use cases,
even though crypto is the great financial primitive for AI.
So that's my take on it.
Love it. Love it. I like it a lot. Mine is as simple as I think crypto will always have something standing on its own. But I do think crypto will, especially with
the agentic commerce piece, I do think that is one of the areas where I think we'll see continued adoption because, well, the unit
economics of AI just are not realistic long-term. Even if we continue to see improvements, it's just
everything is becoming more expensive. And so there's a point where the system cracks and,
well, if people have to start paying for it, it's a lot more feasible for you to pay
as you go instead of 10,000 subscriptions,
which I know is hurting everyone's wallets, including mine.
Yep.
I feel it.
That being said, any final thoughts, James?
This was an awesome, awesome discussion.
Any final thoughts?
Do you want to drop any alpha on what's coming on Trails, Polygon, Sequence?
Yeah.
So, I mean, I mentioned at the beginning of the call, we did a agenda gas rebate program on Polygon.
So X-Fore 2 using USDC on Polygon is extremely cheap.
With our facilitator, we're onboarding more facilitators.
We're dropping an agentic, like an entire platform for building agents.
So that'll be coming out very, very shortly.
And something I did mention actually is you kind of talked a little bit, Sawyer, about
how does the UI kind of like disperse or like should developer platforms kind of build
for an agentic first world?
And so this is something that we're already accelerating where we're making all of our
APIs, not all of our APIs, but a significant portion and the valuable ones X4R2 enabled.
So they're also completely compatible with agents.
So you don't need to go to a developer dashboard and register to get your API key.
Your agent can just do it end to end and paper everything using X4R2.
That's what I meant.
Beautiful sound effects. that's what i meant oh beautiful sound effect so okay i'll stop the shilling but i think that's a good wrap up from my side incredible no i absolutely love it um really really great stuff
uh coming out from from polygon as a whole uh trails and sequence um we've got the the brain
behind trails here with us james thank you so much for joining.
As always, if you're building in X402,
building agents, building anything,
reach out to any of us.
We are here to support and build this with you.
So please don't hesitate to reach out.
If you have any questions on Scale,
you know where to find me.
And of course, if you have any questions
on Polygon, Trails, or Sequence,
make sure to shoot James a DM.
Thank you everybody so much for joining us today.
And we'll be back next week with somebody else who hopefully can carry a conversation like you.
Ciao, ciao. Bye.
Thanks, guys. Bye. Thank you.
