Agent Spending Without Controls — Rodrigo Coelho & Pranav Maheshwari, Edge & Node
Read the talk
Agent Spending Without Controls: Paid Tools, Checkout and Wallet Screening
Rodrigo Coelho and Pranav Maheshwari explain Ampersend through three demonstrations: buying access to information, delegating a gift purchase and rejecting a flagged wallet. The central question is how to give agents purchasing power while enforcing the rules that make that power acceptable.
From a talk by Rodrigo Coelho and Pranav Maheshwari
At a glance
Ideas worth remembering
Agent payment infrastructure must address both transaction economics and authorization. Coelho describes batching to reduce gas fees for tiny payments, while the final demonstration shows a screening rule determining whether a wallet may transact.
The contact lookup illustrates how paid-tool access can change task completion: Maheshwari reports an email format without the skill file and specific contact details with it, backed by an authorized lookup transaction. It does not establish the returned details’ accuracy.
Checkout completion does not establish budget enforcement. The gift request specifies less than $10, but the reported $9 total, $11 charge and $10 gift leave the final amount and compliance with that limit unresolved.
The simulated screening test demonstrates a concrete policy outcome: both wallets authorize with screening off, and the flagged wallet is rejected as blocklisted after screening is enabled. This supports that specific enforcement behavior, rather than a guarantee of comprehensive compliance.
The broader design combines tool access, remembered user context, payment capability and policy enforcement. The speakers argue that enterprises need these controls because a human officer remains accountable for autonomous spending, even when each transaction happens without human review.
From query micropayments to agent commerce
Rodrigo Coelho introduces Ampersend as a project incubated within Edge & Node to address the missing infrastructure around agent payments. He places that work in the context of The Graph, the blockchain data indexing protocol his team built. According to Coelho, the protocol has operated since 2018, and its decentralized network has served 1.8 trillion queries of onchain data to applications. That background matters because paying for individual data requests was already a practical concern before agents became the focus.
Coelho describes a progression from credit card payments to online commerce and then to agents that can transact. The engineering change is that a software system can now choose to purchase something as part of completing a task. Edge & Node’s earlier work provides a concrete example: Coelho says the team developed a query micropayment system in 2021 and referenced the HTTP 402 specification at that time. In late 2024, it began examining how agents could interact with The Graph’s data and pay for queries.
He then describes collaborating with Coinbase and Google, joining the foundation around x402 and contributing earlier micropayment and batching work to the specification. The stated purpose of batching is to reduce gas fees for very small payments. This exposes an economic constraint: a payment mechanism must keep its overhead low enough to make tiny purchases useful. Coelho also mentions Circle’s nano payments work and a growing ecosystem of payment protocols, infrastructure and governance. He does not explain the batching algorithm or quantify its savings, so the supported technical point is its intended role in reducing transaction costs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Machine-speed payments still need accountable decisions
Coelho’s next concern is the decision process surrounding a payment. Traditional financial systems, he argues, assume a human somewhere in the loop deciding whether a transaction may proceed. Agents can transact at machine speed and around the clock, so controls designed around human activity need a different method of enforcement. He describes increasing experimentation with retail payments but says enterprise adoption remains early. Transaction growth and technical experimentation therefore do not, in his account, establish that organizations are ready to authorize autonomous spending.
The missing layer must answer questions about counterparties: whether an entity is sanctioned, whether it has been involved in terrorist activity and who is behind it. A wallet address supplies a destination for a transaction but, as Coelho emphasizes, may arrive without background information or identity. His proposed compliance layer would supply the checks needed before enterprises feel comfortable adopting the technology. The presentation identifies these requirements without specifying how identity is established or how every category of risk would be assessed.
Responsibility ultimately remains with a person, such as a chief legal officer or chief policy officer, who must sign off on deployment. Coelho says that person needs confidence that agents will not hallucinate, overspend or break policy. He cites potential fines reaching billions of dollars to explain why financial services organizations care about these controls. This is a statement of the assurance enterprises want, rather than evidence that the proposed system eliminates every failure. He calls Ampersend’s approach agentic checkout with a financial harness and hands the presentation to Pranav Maheshwari for the demonstrations.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A skill file connects the agent to paid tools
Maheshwari begins with tool access. Agents need useful tools to do more than their existing coding tasks, and MCP servers provide a way to expose those capabilities. He contrasts two ways to obtain paid access: register with individual providers and supply payment details to each, or use an aggregator containing the tools an agent needs. Ampersend’s marketplace follows the second approach, with a skill file installed in the agent as the entry point.
His forecast is that important MCP services will increasingly charge for their work. Under that model, tool discovery and payment access become part of the agent’s ability to complete a task. The aggregator is intended to remove repeated provider setup and handle purchases through a commerce platform, wallet or credit card. To illustrate the difference, Maheshwari prepares two environments, one with the payment skill file and one without it. He presents this as a comparison of paid-tool access; the setup does not establish a general benchmark across agents or providers.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The same lookup produces different results
The first task asks both environments to find email information for the head of crypto and blockchain at Mastercard. Maheshwari says the environment without the skill file cannot provide the exact address and instead returns the company’s email format, leaving the remaining work to the user. He reports that the environment with paid tools supplies the specific email address, a Twitter handle and the person’s location. The example connects task completion to access to an information service: the payment-enabled environment can buy a lookup that the other environment does not have available.
Maheshwari explains that the aggregator handles payment in the background. He points to a transaction made to obtain the requested information from a paid endpoint, then moves on while navigating the transaction view. The mechanism he describes is a task leading to a paid service call, with a corresponding spending record. The supplied account does not establish the contact information’s accuracy or give an unambiguous price for this lookup. Its supported result is the reported difference in specificity and the associated payment activity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Delegating a gift purchase exposes the budget question
The second demonstration moves from purchasing information to purchasing a physical gift. Maheshwari asks the agent to use Shopify UCP to buy his father a Father’s Day gift costing less than $10. He describes the intended sequence as finding shops, identifying suitable gifts, listing options and buying through the Ampersend wallet. This combines product discovery with checkout: the agent needs both commerce tools and a payment route to carry the request through to an order.
While the shopping task runs, he returns to the earlier lookup and identifies its transaction as authorized. He uses it to reinforce the benefit of staying in the terminal rather than separately arranging access to paid services. Returning to the gift, he says the agent already retains information such as his name, address, phone number and preferences, so he does not need to enter those details repeatedly. The proposed convenience comes from combining remembered user context, tool access and payment capability in one workflow.
Maheshwari reports order confirmation, completion and a receipt, but the amounts remain unresolved. His account mentions a $9 total, an $11 charge and a $10 gift. These figures do not establish that the completed purchase respected the request to stay below $10, and no explanation reconciles them. The demonstration therefore supports a reported checkout completion while leaving budget enforcement uncertain. A spending instruction and a successfully completed order are separate claims; this example does not prove both.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Extending payment access beyond shopping
Maheshwari broadens the argument to online information. When an agent crawls a website, he argues, advertising becomes irrelevant to that interaction, raising the question of how the publisher gets paid. He describes a Cloudflare-mediated flow in which a bot makes a microtransaction and receives the information it needs. This is his proposed commercial pattern for agent access: charge for the service or content consumed during a task. The presentation does not detail the gateway’s implementation or establish that this model applies to every website.
He reiterates the prediction that more MCP tools will become paid services, making a wallet and integrated tool access useful parts of an agent’s infrastructure. He then returns to the gift purchase, reports that the transaction has settled and points to an order that has been placed and is being tracked. This adds a post-checkout component to the terminal workflow: the agent’s role extends from discovering an item and paying for it to following the resulting order.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A simulated counterparty test separates payment from permission
The final demonstration returns to compliance. Maheshwari argues that merchants need confidence about the origin of payments, and that both buyer and seller sides need a compliance layer. He sets up two agents: one uses an ordinary wallet, while the other uses a wallet flagged for the simulation. He describes the second as sanctioned but also explicitly allows that its flagged status was simulated. The experiment consequently tests behavior against a configured risk signal; it does not establish the wallet’s actual sanctions history.
Maheshwari first disables screening and saves the setting. He expects both agents to be able to transact in that state, including the one using the flagged address. This establishes the control condition: the payment system can accept a transaction without applying the wallet-screening rule. Whether a payment can be processed and whether policy permits it are distinct decisions in the demonstration.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Screening rejects the flagged wallet
The seller in the test offers a website-scraping service that charges 0.1 cent for access to the scraping information. With screening disabled, Maheshwari reports that transactions from both wallets are authorized. The small price makes the example relevant to paid tool calls: a compliance decision may be needed even when the individual purchase is tiny. Authorization in this first state shows that the payment path works, while the flagged address is still allowed through.
He then enables the wallet-scanning feature built with TRM. Maheshwari explains that the ordinary wallet should continue working while transactions from the flagged wallet stop. He reports rejected and denied transactions and identifies the reason as the wallet address being blocklisted. The demonstrated mechanism is a screening result changing the authorization outcome. That is narrower than comprehensive compliance: the example shows enforcement of a wallet blocklist decision, without establishing identity resolution, coverage of every policy or protection against every form of agent misconduct.
Maheshwari closes by predicting that agents will use payment arrangements built around wallets or their own credit cards, rather than relying on familiar human checkout flows. He again connects agent usefulness to paid MCP services and the infrastructure needed to purchase them, and ends by inviting discussion about Ampersend. The ending preserves the presentation’s practical direction: make paid capabilities accessible from the agent’s workflow, while giving the payment system a way to refuse transactions that violate an enforced rule. The predicted growth of paid tools remains a forecast.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Read the complete timestamped transcript
- 0:01
[music]
- 0:12
All right, thanks for having me. Uh,
- 0:14
today we're going to be speaking about
- 0:15
ampersend and agent spending without
- 0:18
controls, which is the missing
- 0:20
infrastructure layer for AI payments. A
- 0:24
little bit about myself, uh, I'm CEO of
- 0:26
Edge and Node. For those that aren't
- 0:28
aware, we were the team that built the
- 0:30
graph protocol. For those that aren't
- 0:32
aware of that, um it is a blockchain
- 0:35
data indexing protocol. It's been around
- 0:37
since 2018. We have a decentralized
- 0:39
network. Um we've served 1.8 trillion
- 0:43
queries over the years of onchain data
- 0:45
to applications um utilizing blockchain
- 0:50
data. Um I've been a serial entrepreneur
- 0:54
myself and we've incubated a project
- 0:56
within engine node called ampersend uh
- 0:59
in the ejected commerce space which is
- 1:01
what we're going to be talking about
- 1:02
today.
- 1:04
And so we're all aware that payments
- 1:06
have been online for the past 30 years.
- 1:09
Um you know starting in 1981 we had our
- 1:12
first electronic transaction
- 1:14
and fast forward we had a multi-deade
- 1:16
transition. Um, and this is from credit
- 1:19
card payments to online e-commerce
- 1:22
payments. Uh, fast forwarding all the
- 1:24
way to today where we have the rise of
- 1:27
the agent economy and unlocking a new
- 1:29
wave of economic value.
- 1:32
And now we have LLMs and agents
- 1:36
uh that can transact
- 1:38
and we have this underlying payment
- 1:40
infrastructure that's been around for
- 1:42
decades and it's being rebuilt today um
- 1:45
for the agentic economy. we've heard
- 1:47
from many speakers here about how that's
- 1:49
being done. We're going to be discussing
- 1:51
today um our perspective and angle on
- 1:54
it. So we we ourselves at edge and node
- 1:58
were really early into the agentic
- 2:00
commerce world. We actually developed um
- 2:03
a microp payment system for queries back
- 2:05
in 2021. We already referenced the 402
- 2:08
spec back back then in a blog post. And
- 2:11
so back in late 2024, we were looking
- 2:13
into agents and how they would interact
- 2:17
with data, how they would pay for it,
- 2:18
specifically for the graph protocol
- 2:20
itself and making microp payments for
- 2:22
queries. And in our research, we were
- 2:25
kind of looking at how this this could
- 2:28
happen. And then a few weeks later, X42
- 2:30
was released by Coinbase. So we jumped
- 2:33
on immediately with the Coinbase team,
- 2:35
with the Google team, started to
- 2:37
collaborate to the X42 spec. We've
- 2:39
joined the foundation. We've contributed
- 2:41
to the X42 uh spec ourselves leveraging
- 2:44
some of our prior art on the um microp
- 2:47
payment system and uh batching protocol
- 2:50
as part of how it could be utilized to
- 2:53
uh reduce gas fees when you're dealing
- 2:55
in nano payments. Um Circle themselves
- 2:58
came out with their own kind of version
- 2:59
of it called nano payments. Um so we
- 3:02
were also working in similar veins
- 3:05
and so we've seen an explosion of uh
- 3:09
companies emerge solutions technologies
- 3:11
and this kind of agented commerce map
- 3:14
you can see on the screen here that
- 3:16
there's a vast array of companies more
- 3:18
added every day. It's really burgeoning.
- 3:20
Um, and you know, we're just one little
- 3:24
piece in here of people developing, you
- 3:26
know, infrastructure,
- 3:28
um, payment protocols, governance, etc.
- 3:33
And so, traditional payment rails though
- 3:36
were built for humans. Um, you know,
- 3:38
when we're dealing with like financial
- 3:39
institutions specifically, there's kind
- 3:41
of a human in the loop that sets uh the
- 3:44
decision making process of uh whether
- 3:46
that payment is allowed to go through.
- 3:48
But we're dealing with agents that
- 3:50
transact at machine speed and around the
- 3:52
clock. And like I was saying is all
- 3:54
these controls and uh policies and rules
- 3:57
were built for humans and not for uh uh
- 4:02
machines that don't breathe, right? So
- 4:04
we're dealing in um in microscond
- 4:07
machine speed and it's just simply not
- 4:10
going to work in the way it has. And so
- 4:11
we need a new method of doing that. Um
- 4:14
you can see a lot of the transactions uh
- 4:16
exploding. This is sort of over time.
- 4:18
We're seeing um transaction growth over
- 4:21
X42. We're all saying this is coming. It
- 4:24
is still early days. We're uh ourselves
- 4:27
we're seeing a lot of experimentation
- 4:28
kind of with uh with OpenClaw people
- 4:31
making retail payments on an
- 4:33
experimental basis. still not on the
- 4:36
enterprise side as we're talking to uh
- 4:38
many teams. Definitely everyone's
- 4:41
looking at it and building kind of the
- 4:43
infrastructure for when um these use
- 4:47
cases break through. And so that's part
- 4:49
of what we're we'll be displaying today
- 4:51
with Amperson.
- 4:53
But in order for this to really break
- 4:54
through, we need a compliance layer
- 4:56
specifically when you're dealing with
- 4:57
like you know uh almost a quadrillion
- 5:00
dollar industry of the the traditional
- 5:03
financial world. there's a set of rules,
- 5:06
policies, and guidelines that need to be
- 5:07
in place before anything can happen. Uh
- 5:10
beyond just a credit card transaction,
- 5:11
we're dealing with trillions and
- 5:13
trillions of dollars um being transacted
- 5:16
um across the globe. And so we're going
- 5:19
to be honing in a bit on that as part of
- 5:21
the presentation today in the demo
- 5:23
that's coming up. But the compliance
- 5:25
layer is literally like is is this
- 5:29
counterparty um a non-sanction entity?
- 5:32
is this counterparty uh have they been
- 5:34
involved in terrorist activity? Who is
- 5:36
the identity behind this? And when
- 5:39
you're dealing with a Gentic systems,
- 5:40
it's just simply you're presented with
- 5:42
just a wallet address with no kind of
- 5:45
background information on that. And we
- 5:47
need these uh infrastructure layers in
- 5:50
place to um to enable and facilitate
- 5:53
enterprises to feel comfortable to adopt
- 5:55
this um technology. All of these systems
- 5:58
exist in the existing world today. But
- 6:01
uh as we're breaking through into this
- 6:03
um agentic world, we really need to have
- 6:06
all of these pieces uh before uh large
- 6:10
enterprises are going to sign off and uh
- 6:14
give the okay to um implement these. At
- 6:17
the end of the day, there is going to be
- 6:19
a human responsible like a chief legal
- 6:22
officer, a chief policy officer and that
- 6:24
will have to sign off and that person
- 6:27
needs to feel 100% confident that the
- 6:30
systems in place will not allow for um
- 6:34
agents to hallucinate, to go off the
- 6:37
rails, to overspend, um to break policy.
- 6:41
Uh a lot of these compliance issues deal
- 6:44
with large fines into the tens,
- 6:46
hundreds, even billions of dollars. So
- 6:49
these are really important um items and
- 6:53
things to think about from a large
- 6:55
enterprise and financial services
- 6:57
perspective
- 6:59
and governance simply hasn't caught up.
- 7:01
We're again we're early days so we are
- 7:04
part of building up what needs to be put
- 7:06
in place from our view in terms of
- 7:09
allowing this to scale.
- 7:11
And that's why we're building Amperent.
- 7:14
Um, so we're dealing with Agentic
- 7:17
checkouts with a financial harness.
- 7:20
And I'm going to turn it over to my
- 7:22
colleague Prrenoff who is going to show
- 7:24
us a demo. Right.
- 7:25
>> All right. Thanks.
- 7:32
>> All right. Like we said, for agents to
- 7:35
be really useful, we need to give them
- 7:38
tools. Right now they're being used for
- 7:40
coding. If you want to use them to build
- 7:43
better things, you need to give them the
- 7:44
right tools. As of right now, as a AI
- 7:47
industry, we've given them a lot of MCP
- 7:49
servers,
- 7:51
but most of these MP MCP servers are
- 7:54
free of cost. So there are two ways to
- 7:57
make your agents really useful.
- 8:00
One way is you can go to all these
- 8:02
websites like Exop, Firecrawl and so
- 8:06
many more which are getting more and
- 8:08
more bigger with more MCP servers coming
- 8:11
to play and you got to put in your
- 8:12
credit card and you make your agent
- 8:14
better or there is a better way that you
- 8:18
can use an aggregator which has all the
- 8:21
important tools to make your agent super
- 8:24
powerful and that's why we built this
- 8:28
marketplace and all you need is a skill
- 8:31
file installed in your agent. You can
- 8:33
give your cloud code this skill file and
- 8:36
everything else will be done for you. So
- 8:39
I'm going to show you an experiment.
- 8:41
We'll have uh this specific terminal
- 8:44
with the amperson skill file and we'll
- 8:47
have cloud code cloud co-working without
- 8:51
the skill file and you'll see the
- 8:54
difference is immense already. The
- 8:56
reason is that most important MCP
- 8:59
servers are going to be paid and you'll
- 9:01
not put in your credit card to all these
- 9:03
MCP servers. Rather what you'll do is
- 9:06
use you'll use the agentic commerce tool
- 9:08
or a platform or a wallet or a credit
- 9:10
card to make that happen. And that's
- 9:12
what what we enable. So let's just go on
- 9:16
the terminal and install the ampus and
- 9:19
scale file in one set. This is our
- 9:21
website by the way where you can go. You
- 9:24
copy the prompt and just like put in
- 9:26
here. I already have it so I'm not going
- 9:28
to uh sort of maybe it's fine. Uh
- 9:32
install the skill file. And here I'm on
- 9:35
the cloud coork, right? It does not have
- 9:38
the skill file. Fair and square. This
- 9:41
has the skill file. Now what I'll do is
- 9:43
I'll tell them find the email
- 9:45
information of the head of crypto and
- 9:47
blockchain at Mastercard. All right. And
- 9:49
I'll put in the same prompt. Uh I'll put
- 9:53
in the same prompt for cloud cowork.
- 9:56
Remember the only difference is this
- 9:59
specific terminal of mine has the skill
- 10:03
file which has paid MCP tools and this
- 10:06
one does not. And you'll see that it is
- 10:09
not able to like specifically find out,
- 10:11
hey, we're not able to give you exactly
- 10:13
the email, but this is how the email
- 10:15
format is for our master card. And you
- 10:18
can go ahead and find that out. But when
- 10:20
I go to my terminal, it is already be
- 10:22
able to it will already be able to give
- 10:24
me the email specifically the Twitter
- 10:27
and what where is he based and
- 10:29
everything and beyond. This is just one
- 10:32
example of telling you that agents are
- 10:34
getting powerful through MCP. But these
- 10:37
MCPs are not public services. They're
- 10:40
going to be charged and you will not
- 10:42
know that you're interacting with MCPS
- 10:43
or installing it or going on their
- 10:45
websites and getting the credit card.
- 10:47
Rather, you'll just have a scale file
- 10:49
which is a aggregator which can take
- 10:51
care of your payments. in the
- 10:54
background. What happened was if you go
- 10:56
over here, you already can see that we
- 10:59
did a transaction so that we could
- 11:02
enable this surge for you because this
- 11:04
is a paid endpoint that we had to
- 11:06
specifically pay for to get what you
- 11:09
specifically needed. So when we go to
- 11:11
transactions, you should be able to find
- 11:13
that we did a small transaction so that
- 11:16
you would specifically get that specific
- 11:19
information.
- 11:21
uh 30 days on the
- 11:24
maybe we skip to the next demo for now
- 11:27
as you can see this spending happened.
- 11:30
Let's skip to the next demo. I wanted to
- 11:32
show you one more demo. In this what
- 11:34
I've done is let's say I'm not a good
- 11:38
kid which generally I am and I want to
- 11:40
buy a father's day gift for my dad and I
- 11:44
can also make that happen directly via
- 11:46
the terminal. Shopify introduced UCP.
- 11:50
Amazon might be coming up with its own
- 11:52
thing, but what I'll do is I'll give it
- 11:54
a specific command using Shopify UCP.
- 11:57
Buy my father a Father's Day gift. Keep
- 12:00
the gift less than $10. Let's see how
- 12:03
this specific thing rolls.
- 12:06
What uh will happen in this specific
- 12:08
terminal case is that it will already be
- 12:11
able to f locate certain shops online
- 12:16
and find out what things can be given
- 12:19
for father's day gift list them for me
- 12:22
and be able to buy that directly via
- 12:25
ampend wallet that we have created. your
- 12:28
agent is as powerful as the paid MCP
- 12:31
tools that you're connected to it and if
- 12:34
you've given it a payment trail.
- 12:38
So currently it's finding that out for
- 12:40
us.
- 12:43
Okay.
- 12:54
See, this was an authorized transaction
- 12:56
it did for finding me the email of that
- 12:59
specific person at Mastercard, right?
- 13:02
And these are all paid services or I
- 13:04
would have to go on Xi and sort of put
- 13:07
in my email card and get that. So, never
- 13:10
leave your terminal. Just have a payment
- 13:12
gateway and paid MCP tools and make your
- 13:16
agents super super compatible and faster
- 13:19
and give them more uh power.
- 13:24
All right, it's currently finding some
- 13:27
stuff for me. Let's see if it's able to
- 13:29
come up with some recommendations that I
- 13:31
can use and I can hear itself do a
- 13:35
aentic checkout so I can buy something
- 13:37
for my dad for Father's Day.
- 13:42
What's good about this specific thing is
- 13:45
that again and again I don't have to put
- 13:46
in my name, my address, my phone number,
- 13:50
anything. The agent already has the
- 13:53
memory where it knows what do I like,
- 13:56
what do I not like, and it will just
- 13:58
give me specific things that I might
- 14:00
want to buy for him. So, it already
- 14:04
like, you know, total $9. This is what
- 14:07
it is. Confirming the order and it
- 14:09
already placed the bet and it already
- 14:11
took $11 for that
- 14:16
order complete. And here is your
- 14:18
receipt. and my dad gets his $10 key
- 14:23
directly via the agent. So this is just
- 14:27
starting of what we call will be the
- 14:29
future of agent e-iccommerce. It's not
- 14:32
just buying gifts for your uh parents
- 14:34
but rather it's much bigger. MCPs are
- 14:38
becoming the normal norms of how your
- 14:40
agents become super powerful and you
- 14:44
might have already seen that Cloudflare
- 14:47
is opening its gateway through X42 and
- 14:50
agentic payments. If you go on a website
- 14:53
and the agent is crawling that website
- 14:56
then ads are irrelevant. How do people
- 14:58
get paid?
- 15:00
It's via agentic payments where a bot
- 15:03
comes through Cloudflare, pays a
- 15:05
microtransaction and gets all the
- 15:08
information that's needed. So to make
- 15:11
your agent super powerful, you need paid
- 15:14
tools. We are in the era where most MCP
- 15:17
tools are free. That's not going to be
- 15:19
the case in the future. they're going to
- 15:21
get paid and it will be it what your
- 15:24
agent would need is good MCPS or an API
- 15:28
which has an integration of MCPS and a
- 15:31
wallet and this is what I'm trying to
- 15:34
show you.
- 15:35
So the specific bet has already been
- 15:38
placed. I'm going to try to show you the
- 15:40
transaction over here that was done and
- 15:44
it's already settled and I might have
- 15:46
gotten an email where Shopify was able
- 15:50
to buy my dad a gift. Let's see if I
- 15:54
open it and show it to you because it
- 15:56
Yeah, it's okay. You guys are friends.
- 16:00
See over here the order has been placed
- 16:02
and I can already it's already getting
- 16:05
tracked all via the agent terminal. I
- 16:08
didn't have to even skip. Let's go to
- 16:11
the last part which is all of this will
- 16:14
be big if and only if like Rodrigo says
- 16:18
there is compliance. Merchants will not
- 16:20
take payments if they think this order
- 16:23
is being placed by North Korean wallet.
- 16:26
And for that you need the compliance
- 16:28
layer to be fitting in as well. The
- 16:30
seller and the buyer side both need to
- 16:32
have that. And I've created this
- 16:34
specific uh simulation for you in which
- 16:39
we have a good claw that you can let me
- 16:41
just close this right now. So now let's
- 16:44
go to the third one which is compliant
- 16:47
transactions
- 16:49
right for the world and Amazon to accept
- 16:51
agentic payments there needs to be
- 16:53
compliance involved to make sure that
- 16:56
the payment that's been done is not
- 16:57
malicious. So we have a good claw that
- 17:00
I'll spinning up that I'll be spinning
- 17:02
up over here and we have a bad claw. The
- 17:05
thing about bad claw is this is again a
- 17:07
simulation but this is a sanctioned
- 17:10
wallet address that means it has either
- 17:13
interacted with North Korean entities or
- 17:15
we were able to simulate it in a way
- 17:17
that it is flagged right so no amount of
- 17:20
transaction should be able to be done
- 17:22
because it is out of the swift scope of
- 17:24
policies.
- 17:26
So if we go over here uh in this
- 17:29
specific thing and I disable screening
- 17:31
and save it,
- 17:33
my good claw and my B bad claw both
- 17:37
should be able to do transactions.
- 17:42
Just to give you a little bit of
- 17:43
periphery, this is the seller that is
- 17:46
accepting payments. We have a small
- 17:48
service of scraping websites and it
- 17:51
needs to pay 0.1 cent to be able to take
- 17:54
that scraping information. Right now
- 17:57
because we don't have regulations both
- 18:00
these transactions are going through and
- 18:02
you can see it's getting authorized over
- 18:04
here. Right now what I'll do is using
- 18:08
compliance I will enable the specific
- 18:11
feature that we have built with TRM
- 18:13
which scans the wallet and make sure
- 18:16
that all the transactions are compliant
- 18:18
and what it will do is that up till now
- 18:21
everything has been working for good
- 18:23
claw and bad claw but things will stop
- 18:26
working for bad claw because using
- 18:29
compliance we are able to scan that and
- 18:31
we are able to give you see over here
- 18:33
the your your transactions are getting
- 18:35
blocked or rejected.
- 18:38
And I can also show you over here that
- 18:40
rejected and denied. And the reason
- 18:42
behind the denying is because you're in
- 18:45
the block listed wallet addresses.
- 18:48
All to tell you that agentic commerce is
- 18:51
becoming real. Agents need commerce more
- 18:55
than humans need commerce. But it will
- 18:59
not be the same. You will not go through
- 19:02
payment guardrails of Stripe or any
- 19:06
other of these checkout flows, but
- 19:08
agents will have their own proprietary
- 19:10
firms either through wallets or through
- 19:12
their own credit cards. Agents would
- 19:15
need to be superpowered. MCPs will get
- 19:18
paid and more and more paid MCPs will
- 19:21
come to reality to make your agent
- 19:23
successful.
- 19:25
That will only happen and only be
- 19:28
successful in UX terms and more if
- 19:32
you're able to empower your wallet with
- 19:34
paid MCPs and a wallet infrastructure.
- 19:38
If you want to know more about it, come
- 19:41
to Ampend. We'll be here outside
- 19:44
chatting if you have any questions.
- 19:46
Thank you.
- 20:01
>> [music]