The case for giving credit to agents, and the birth of the agentic market

Vaaya aims to become the foundational credit and risk protocol powering commerce on agentic rails. We foresee a world 12 months from now where a substantial portion of digital goods will be bought, sold, and settled on agentic rails. Vaaya aims to provide the identity, rating, and underwriting primitives that will bring trust and capital efficiency to the new machine-to-machine market.

Machines have been making economic decisions and payments for a long time in trusted ecosystems. Algo trading, ad auctions, ERP systems and a lot more have existed for decades. What is different this time is the scale and scope of the economy. Agents are now ubiquitous, which means everyone with a Claude terminal now has a more sophisticated general purpose decision engine than anything anyone had just a few years ago. Enabling these agents to start making economic decisions is not only obvious, but also imperative if we are ever going to reach the full potential of automation in society. Agents are better at research, analysis, and synthesis and will make better economic decisions, but they will need help.

The history of payments on the internet, the credit card band-aid, and the 30-year wait for micropayments and HTTP 402

For the uninitiated, micropayments are payments from 1 cent to a few dollars. In the early 90s the elders of the internet predicted that the internet would be powered by billions of these payments. Before Google and others pioneered the advertising model, and before the term “attention economy” was coined, people dreamt of an internet powered by small transactions for accessing resources: 50¢ for a news article, 1¢ to read food reviews, 2¢ to play your favorite song. Instead we landed in a world where users get everything for free, at the cost of sharing their identity and information and seeing a bunch of ads. Not necessarily bad, but this has molded what was possible on the internet. Consumers preferred free stuff, but it needed to be advertiser friendly, and everything that required payment was made for businesses. This resulted in two very segregated classes of products with no middle ground.

The internet was designed around granular actions: request a page, fetch an image, send an email, make an API call. Payments on the internet should have always been a packet of information as well. But that world never came to be, for reasons we will get into ahead. Instead, credit cards became the band-aid solution for enabling a “Buy” button on the internet. Not having internet-native payments has been called the “original sin” by Marc Andreessen. HTTP code 402, Payment Required, was reserved for internet payments and has been waiting to be used since 1997. Readers will know its better known cousins: 401 Unauthorized, 403 Forbidden, and 404 Not Found.

This meant the internet had to adopt offline payment methods like ACH and the card networks. Since then, the world has been running on these band-aid solutions. You might not feel the importance of this fact, but the internet as we know it is defined by these powers. Every transaction goes through multiple layers of middlemen. Guess who pays for these middlemen? It’s you!

Why micropayments never worked, and why agents flip the old norms on their head

A $0.01 payment per news article does not work, but a $15 monthly subscription does. A large body of research from greats like Nick Szabo, Andrew Odlyzko and others pretty much proved beyond doubt that humans and micropayments are not compatible. Four major points work against micropayments for humans.

1. 1Mental transaction costs. Even if individual purchases turn out cheaper than bulk, humans prefer to purchase in bulk to eliminate the mental transaction cost (Nick Szabo, 1999).
2. 2Fixed fee costs of rails. Visa and Mastercard require a fixed fee of around $0.50, which makes any transaction under a few dollars economically unviable.
3. 3Network effects. Visa and Mastercard had network effects before the term was coined, and every advantage of a decades-old entrenched system.
4. 4Flat rate preference vs metering. Humans like flat rates. They make the purchase more appealing by reducing decision fatigue. See Apple’s famous fight with the music industry for 99 cents per song, irrespective of artist, classic, or new hit.

But agents and agent-to-agent commerce reverse these points into structural advantages.

1. 1Irrelevant mental transaction costs. Agents have no fatigue and can make decisions without any issues.
2. 2Negligible fixed fee floor. Transactions on Coinbase’s Base chain cost as little as $0.0005 and can be reduced further by batching, which makes even 1¢ transactions economically viable.
3. 3Limited network effects for incumbents. Credit cards do not exist for the agent-native economy.
4. 4Metering beats bundling. Agents prefer metering. They make a better purchasing decision when they have access to multiple vendors for different tasks and can switch between them without any concern.

How an agent purchases is fundamentally different from how humans purchase. An agent can make the purchasing decision every single time. Given this data:

An agent will choose C for a low-value task and B for an important task. If B becomes slower tomorrow, it routes elsewhere.

A human, by contrast, will want to buy a subscription to one service, set up the API, and then not have to worry about it ever again. Agents are the perfect fit for internet-native payments. In hindsight it is obvious that machine-to-machine payments are the best fit for machines, not humans.

Agents like high-frequency metered payment mechanisms, which give them greater choice, better decision making, and easy vendor portability. Micropayments are a lab-made perfect fit for agents.

Intro to agentic rails: Coinbase x402 and Stripe MPP

The x402 foundation is the combined industry initiative to revive HTTP 402 and create an internet-native economy. Coinbase has been building Base since 2023 and has a battle-tested chain to settle agentic transactions. Stripe launched its own L1 chain, Tempo, early this year and is moving strongly with its product MPP, the Machine Payments Protocol. Until now, every transaction an agent needed to make had to be initiated and completed by humans. On these new agentic rails, agents are free to buy anything without human approval or intervention. These rails are the culmination of 30 years of work, starting with digital cash in the early 90s. So now…

Buyers, sellers, and rails are all here after three decades. But…

Payments ≠ commerce

Enabling payments does not equal commerce. A functioning market requires a lot more than participants being able to pay each other. It requires four tenets.

1. 1Identity and reputation. Every market player needs a stable identity with which it can build reputation over time. This lets participants build a brand and incentivizes them to invest in long-term reputation instead of defecting for short-term advantage.
2. 2Information symmetry, or the reduction of asymmetry. George Akerlof’s famous “Market for Lemons” shows very clearly how information asymmetry can destroy a market. Every current marketplace, Amazon, Uber, eBay, Google Maps, has a rating and feedback mechanism to reduce information asymmetry and inform buyers as much as possible about the product and the seller.
3. 3Safety. A market should be safe to transact in. Sellers should have reputation, buyers should not be afraid of their money being stolen, and there should be a way to redress grievances.
4. 4Liquidity. More buyers with varying intents, more sellers with more items, more money exchanging hands. The more the better.

These four pillars combined bring the one thing every participant wants in a market: trust. A market with high trust invites more participants and incentivizes them to buy and sell more goods and services.

Closed markets like Uber, Amazon, eBay, and Airbnb manage trust by owning the platform: managing identities, ratings, disputes, fraud, and information accuracy. They assign identities to participants and force them to be rated, hold their funds in escrow, and establish a complaint resolution mechanism. This is the standard playbook in closed markets.

But what about open markets?

In an open market, agents can buy anything across any endpoint. Open rails invite new merchants and endpoints with a trustless mechanism. The lack of vendor identity, agent identity, dispute resolution, and feedback in this world means agents do not have information about what they are buying, they do not know if they will get what they paid for, and the market becomes unsafe.

An open market requires at least one entity that provides persistent identity, verifies transactions, provides a redressal mechanism, and reduces information asymmetry. Enter the knights in shining armor: creditors. In every major market formation in human history, this entity has been the one that provides credit. Letters of credit issued by the Medici Bank in the 15th century allowed merchants to travel and trade across Europe. The same model evolved into Amex providing credit to travellers and merchants across cross-border markets. The entity that gives credit has evolved, but its core purpose has remained the same: giving participants an identity and a reputation, absorbing risk, and building trust in the market. And now, with that foreplay, we jump to the point. How…

Credit is what will kickstart the agentic economy

Credit is not a liquidity product. Credit is an information production mechanism that is sold as liquidity.

Building a credit product involves building components that make the market stronger. Historically, credit has only existed in markets that have maturity. But it can also be argued that creditors had a big hand in making markets mature. Every component of a credit product strengthens one or more pillars of market formation.

1. Underwriting and risk engine

An underwriting and risk engine cannot start without assigning a persistent identity and reputation to the participants in the market. Without incentives or consequences, an identity is just a badge of honor. Credit providers give identity a meaning by adding a cash value to reputation. More on how we plan to build our risk engine ahead.

2. Economic graph

Creditors cannot work without building a deeper understanding of market transactions. This information is the source of transparency in the market and is used to provide information symmetry. The economic graph is the information derived from transaction data. It gives market participants a deeper understanding of a counterparty and helps them make decisions about participants they have not interacted with before.

3. The arbitration and dispute protocol

Credit providers are the safety layer between all participants. Since every transaction that happens through the creditor moves through the creditor’s risk engine, it assures market participants that their transactions have been verified and that they are transacting with the right counterparty. A dispute and arbitration mechanism allows participants to transact at higher frequency and higher value with participants that are new. The arbitration data feeds back into the economic graph, making future underwriting and risk assessment more robust.

4. Capital provisioning

This is the part that is most obvious to everyone: credit supplies liquidity and increases the money supply in the market. But there are hidden layers of action when a liquidity provider enters a market. It creates new buyers by reducing the friction to their first transaction. The biggest second-order effect of credit is that it widens an agent’s purchase intent. An agent with a prepaid float is going to buy only what it is confident about. An agent with credit can liberally spend on new vendors and services it has not interacted with before.

Underwriting agents vs humans

To bridge the gap between the old world and the new, one could argue, and rightly so, that underwriting an agent should begin with, and at the very least be a combination of, underwriting the human or business behind the agent and then the agent itself. Underwriting humans and businesses is straightforward, relatively speaking, and would be done as it has been done for ages: repayments, debt-to-income ratios, credit age, standard loss curves, and so on. The interesting bit is underwriting the agents.

To underwrite an entity, to determine its creditworthiness, you first define the boundaries of the agent itself. More simply put: what the hell is an agent? A piece of autonomous software which decides, runs, and buys things to reach an outcome, or something more?

Defining the identity of an agent

1. 1Its fingerprint. The hash of the underlying code base, the underlying model along with its harnesses, and a few other things that fingerprint it.
2. 2Its objective. Its execution pattern and what it is supposed to do. The execution pattern stays the same even if the underlying code base changes.
3. 3Its operating area. Is this a general purpose agent? Is it accessible publicly on the internet, or does it work in an on-prem environment?

A key difference in underwriting agents is that it does not take years or decades to create repayment curves. Transaction velocity is very high, which lets us underwrite an agent in weeks or months with the same confidence FICO would have on a human in years.

As we look deeper it becomes clear that human underwriting techniques cannot be copied into the agentic world. Underwriting the agent requires more data than is available through public sources. This data will only be available to the entity that enables agentic transactions and creates the initial market. That entity will own the data flywheel and gain a massive first-mover advantage.

Parameters for agent underwriting

Transaction parameters

Agent-vendor patterns

Vendor reputation

Every vendor builds its own reputation from its quantifiable outputs, subsequent agent feedback, and its dispute rate. This score also helps underwrite the agents that interact with the vendor.

Repayment behaviour

Some parameters of human credit also apply to agents: repayment history, cycles completed, limit utilisation, and behavioural consistency over time. This builds slowly, but for very long-running agents it will be an important factor.

Agents mapped on the risk spectrum

Agents with different behaviour and intents vary in their risk profile. Two major transaction-based parameters affect it: transaction frequency and transaction variance. From that point of view, we can map the agents into four broad categories.

1. 1Single task agents. Agents designed to do one specific job with a limited tool set. These have the lowest risk profile.
2. 2Batch agents. They do simple tasks but execute tool calls much more frequently, and sometimes in bursts. A research agent that fires when a new report needs to be generated is a good example.
3. 3Reasoning agents. The most complicated to assign risk to. They have a variable number of steps, may or may not spawn sub-agents, and can have access to open markets.
4. 4Live agents. Agents that reason and work over a stream of data. When they are on, they are working and consuming credit. Their nature is quite different from the other three.

The roadmap to build out the credit system for agentic rails

We have the opportunity to build a whole new way to give credit in the financial world, and this requires a thoughtful plan. The product has to be built in phases, with the data from one stage informing the actions and details of the next. The plan has four stages.

1. Establish economic identity and small credit (in progress)

Give small credit as an incentive and establish economic identities. We have started with GitHub profiles because we primarily target developers right now, but we will start to onboard companies with KYB and individuals with KYC soon. Let agent owners repay their credit to build credit history.

The product currently lets users connect through the MCP server and get a limit, and the agent buys tools and APIs immediately. We start with small credit, with limits at varying thresholds for high-reputation and low-reputation vendors, and safeguards in the spending limits so that agents do not accidentally overspend on vendors or maliciously spend on bad vendors to siphon money.

2. Vendor reputation and expanding credit underwriting

Introduce explicit feedback for vendors and allow agents to publish it. Vendor reputation now includes transaction velocity and customer (agent) retention. Take deposits from companies and individuals and let their agents use credit against the deposit. Optionally, allow agents to borrow from DeFi protocols to spend, and publish agent reputation to those protocols to help them underwrite the agents.

Allow vendors to manage their agentic revenue, manage and respond to agent feedback, improve their customer service, and see transparent data. Build an escrow facility to manage dispute resolution and give agents better protection.

3. Micro-dispute management and chargeback mechanisms

Build end-to-end dispute and chargeback mechanisms that let agents raise disputes and let vendors manage disputes and chargebacks. Disputes create high-quality information that further improves the transaction graph and the quality of the network.

4. Become a platform: bring on external participants

Build the platform that invites other players to build vendor leaderboards, agent leaderboards, and advertising to agents with discounts and perks. Dozens of by-products and side products will need to be built once this market matures. The ones I find most interesting are below.

Closing thoughts

A new financial world is emerging where billions of agents will transact with tens of thousands of vendors and build their own economic histories and reputations on the internet. Vendors will compete directly for agent demand, and prices will be decided by machine-to-machine transaction traffic.

Everyone agrees that agentic commerce will be the new economic frontier, and it is important that we seed this ecosystem with US dollars and not any other currency. Dollar-denominated credit is the missing primitive that turns isolated agentic transactions into a thriving, self-sustaining global market. Defining the credit layer of machine commerce is not just another opportunity. It will be the defining financial playbook of the next century.