Your AI agent should pay its own bills

The first time you build an agent that can spend money, two things happen at once. The first is small: it stops asking for permission every two minutes. The second is bigger: it stops being yours in a useful sense and starts being itself — a thing that goes out into the world, swipes, retries, picks up a tool it needs, and comes back with the answer.

That second thing is what we are actually selling. Not the wallet. The wallet is a means. The thing we are selling is autonomy with a hard balance attached, and it is the difference between a chatbot and an employee.

What a flat-rate subscription actually charges you for

Pull up any AI tool that costs $20 a month and look at how you used it last week. Three days you didn't open it at all. One day you ran a long agent loop that, if it had been priced honestly, would have cost six dollars in inference. The platform absorbed it because eight other users were paying $20 to watch a cursor blink.

A subscription is a pooled-risk product. Heavy users subsidized by light users. That works for Netflix because watching is bounded — you can only watch so many episodes a day, and the cost to deliver them is rounding-error cents on a $20 bill. It does not work for autonomous agents, where:

• one user spawning a research loop can cost more in two hours than the user pays in a month;
• the platform's only defenses are rate limits, which fire at the worst possible moment;
• the sustainable equilibrium is everyone gets degraded until the price feels fair to the platform.

You have felt this. The model that was sharp in January feels dull in April. The agent that finished tasks last week now stops halfway and tells you to "continue in a new chat." That is not a bug. That is the math working as intended.

What a wallet changes

Give the agent its own wallet — not a corporate card you have to babysit, but a real, tiny, on-chain balance it controls — and the entire negotiation flips.

You stop paying for time access to a service. You start paying for outcomes that were measurably worth their cost. The agent stops competing with other users for a shared rate-limit pool. It pays its own way, exactly to the cent, every call. The provider stops needing to soften every model and can finally serve the heavy task without resentment.

We built Franklin to make this concrete. Franklin holds a USDC balance — typically $5, $20, $100, however much you fund it — and spends from it autonomously across 55+ models, image generation, web search, trading data, and tools. Every single call is a x402 micropayment settled in USDC. There is no API key. There is no monthly minimum. The agent runs until the wallet runs out, then it stops and asks if you'd like to top it up. That's the whole thing.

"Pay per call" is not the pitch

The tech-correct version of this story is "pay-per-call APIs settled in stablecoins via x402." That story is true and it does not sell.

The real pitch — the one that gets a shrug followed by "wait, actually" — is shorter:

Don't worry about credits. Load $20. Go to work. Franklin tells you when it needs more.

This is the same trick that public-key cryptography pulled fifty years ago. The mathematicians had RSA and Diffie-Hellman, and what they sold to the world was a green padlock in the address bar. Nobody buys public-key cryptography. People buy "this site is safe." The thing that made Stripe a generational company wasn't its API. It was seven lines of code and the words "we'll handle the rest."

The wallet is the green padlock. What we sell is "go to work."

What you do with this in practice

The agent-with-a-wallet pattern unlocks four things that flat-rate models fundamentally cannot:

1. The agent picks the right tool, including the expensive one

When the next call costs real money, an agent has to think about whether the call is worth the money. Franklin's smart router does this every request — it looks at the prompt, picks the cheapest model that can plausibly answer, and only escalates to the frontier when the cheap one would fail. We track the savings vs. always-frontier; on a typical week of agentic workflow it's 60–80%.

You couldn't do this on a subscription. The whole point of a subscription is that the price is the same regardless. There's no signal to economize on. So agents on flat-rate platforms either default to the cheapest model (and you wonder why the answers feel thin) or default to the most expensive (and the platform throttles you halfway through).

2. Long-running loops that don't trip rate limits

Autonomous loops — the kind that scrape, summarize, retry, branch — are the highest-value thing you can do with a model and the worst-fit thing for a flat-rate plan. Two hours into a research loop you'd otherwise be cut off. With a wallet you just run.

3. Buying data, not just inference

Once you can pay, the universe of things you can buy expands past LLM tokens. Real-time market data, on-chain analytics, image generation, video, web search. Franklin treats all of these as the same primitive: a tool with a price tag. The agent calls them when they're useful. You see the line item.

4. Spend-per-task accounting that actually means something

You ran the agent. It cost $0.43. You can answer "was that worth it?" There is no smearing across a calendar month, no wondering if you "got your money's worth," no end-of-month surprise. This sounds boring. It is the single most demanded feature by anyone who has ever tried to deploy an agent inside a company that has a finance department.

"What if I don't want to think about money"

Then you don't. You load $20 once a month, the agent works, and you don't think about it. The wallet does not make you care about cost; it lets you care about cost when you want to. That is a strictly larger product than a subscription, which only lets you not-care.

The other version — agents with API keys — forces you to care about the wrong thing. You care about which provider has which key, which key is in which env file, which key has been rotated, which dashboard to log into to top up which prepaid balance. With one wallet that holds USDC, you have one number to watch and one number to refill.

The invariant we are betting on

Flat-rate AI is a transitional product. It exists because crypto rails were not ready three years ago and credit cards were too coarse to settle a 0.001¢ inference call. They are ready now. x402 settles cents — and tenths of cents — natively, on-chain, with no chargeback risk and no platform middle-man rationing.

When the rails work, every agent eventually ends up holding its own balance. Not because crypto is cool, not because it is ideologically correct, but because flat-rate-rationed agents lose to balance-backed agents on quality, on autonomy, and on honesty about what you are buying.

The wallet is not the product. The product is the agent, free to do its job. The wallet is what makes the agent finally able to.

If you want to try it, installation is two commands and the wallet generates itself. Fund it with $5 to see what an agent feels like when nobody is rate-limiting it. That is usually all it takes.