# Can AI agents have credit?

_By Nakul Kelkar, July 20, 2026_

The word has two meanings and they are pulling in opposite directions. When a developer tool says "credits" it means prepaid usage units: you paid in advance, the tool metered your consumption in its own currency, and when the units run out you buy more. That is a coupon with a balance. When a bank says "credit" it means money extended now, on trust, to be settled later. That is a loan, and trust is the whole of it. The question "can AI agents have credit" is only interesting in the second sense, and in the second sense the answer is: not directly, and yes, through the person behind them. The mechanism by which that works is the mechanism we built Vaaya around, so this post is partly a definition and partly an explanation of why we think it is the primitive the agent economy has been missing.

## Why an agent cannot be underwritten on its own

Underwriting is the act of deciding how much to lend to someone on the basis of what you know about them. For a person that knowledge is repayment history, income, tenure, and a set of loss curves that have been refined over decades. For an agent, in 2026, there is almost none of it. An agent has no persistent identity that survives a redeploy, no history that was not produced in the last few weeks, and no recourse: you cannot garnish a language model's wages. A lender who extended credit to an agent as such would be lending to a process that could be replaced by a different process tomorrow, running on a different model, with the same name.

So credit cannot start with the agent. It has to start with the principal, the human or the company that provisioned the agent and is responsible for what it does. That is not a compromise; it is how every credit system in history has begun. The Medici did not underwrite the cargo. They underwrote the merchant, and the letter of credit let the cargo move. The agent is the cargo.

## What can be underwritten

The interesting question is what a lender can know about a developer that a bank cannot. A bank knows a person's financial past. It knows nothing about their work. But a developer's work is, unusually, public, timestamped, and hard to fake at scale. A GitHub profile is years of commits, reviews, releases, issues closed, and repositories maintained, each one an act that someone else could inspect. It is a long record of showing up, and showing up is most of what a lender wants to know.

That is what Vaaya underwrites. A developer connects their GitHub account, and the account receives a credit line sized from that public history. The line is spendable once a card is on file, because the card is what settles it: at the end of each week, the card is charged for what the developer's agents actually drew. The developer never deposits anything in advance. Their agents run from the first day, inside a policy, against a line that exists because of work the developer already did.

The line is separate from two other pots that live on the same account. There is a welcome credit of up to $10 that every account gets and that needs no card at all, so an agent can make its first calls immediately. And there are prepaid packs, $10, $30, or $100, for anyone who would rather pay in advance and keep the whole thing card-free. All three show as one balance to the agent. The agent does not know or care which pot it is drawing from. It sees a ceiling and a policy.

## What the agent gets, and what the developer gets

The agent gets the ability to spend before value has been proven. That sentence sounds small and is the entire point. An agent with only a prepaid balance spends defensively; it buys what it is sure of, from providers it already knows, because every cent it spends is a cent that was deposited. An agent with a line spends the way a funded business spends: it tries the new provider, runs the wider search, generates the second draft. The purchase intent widens, and a wider purchase intent is what turns a catalog into a market. Credit does not just add liquidity to the agent. It changes what the agent is willing to buy.

The developer gets an agent that works on day one, a card that is charged for what was used rather than for a subscription that was mostly not, and a line that grows as settlement proves reliable. None of the controls are loosened by the existence of credit. The key still carries its categories and monthly limit, the agent still states a ceiling on every call, the account still has a ceiling on what it can owe. The line sets how much can be drawn. The policy sets how.

## Why credit rather than prepaid

Prepaid is easier to build and easier to explain, and almost every payment product for agents is prepaid for that reason. It asks for money before value is proven, and it asks nothing else. That is precisely its weakness. A prepaid balance produces no information about the person behind it beyond the fact that they had the money. It cannot grow on its own, it cannot distinguish a reliable principal from a careless one, and it teaches the agent to be timid.

Credit asks for identity instead of money, and identity is generative. **Credit is not a liquidity product. Credit is an information production mechanism that is sold as liquidity.** Every draw against the line, every weekly settlement, every refused call is a fact about the principal and the agent that did not exist before. Over weeks that record becomes an economic history for the agent itself, which is the thing that was missing at the top of this post. An agent cannot be underwritten today because it has no history. Give it a line through its principal, let it transact, and in a few months it has one. Transaction velocity is high enough that an agent can accumulate in weeks the kind of record a person accumulates in years.

That is the long game, and it is why we started with credit rather than a balance. The near-term product is a developer's agents spending against a line their GitHub earned. The thing being built underneath is the history that will let the agents themselves, eventually, be the ones underwritten.

The full argument is in the memo on [why credit](/why-credit), and the account itself is described on the [credit account](/credit-account) page.

## Questions

**Can an AI agent have a credit line?**

Not in its own name. Credit attaches to the person or company responsible for the agent, and the agent draws on it through a key with a policy. Vaaya sizes that line from the developer's public GitHub history and lets their agents spend against it before any money is deposited.

**What is the difference between AI credits and agent credit?**

AI credits are prepaid usage units inside one tool, spent on that tool's models. Agent credit is money extended now and settled later, spendable across providers, and underwritten on the trust the lender has in the principal. One is a coupon. The other is a loan.

**How is an agent's credit line settled?**

A card on the account settles what the agents drew, at the end of each week. The line becomes spendable once the card is on file. Reliable settlement is what grows it.