Para Agent Wallets: $200 a Week for Chipotle

Para agent wallets may be the clearest explanation yet of where AI-powered payments are actually heading. A $200-per-week stablecoin card that can only buy Chipotle sounds like a thought experiment, but Para CEO Nitya Subramanian uses it as the most direct way to explain what her company builds and why she thinks the entire protocol debate has been missing the point.

The Chipotle Card Is Not a Gimmick

Picture an AI agent with a card in its digital wallet. The card draws from a stablecoin balance, has a $200 weekly ceiling, and is restricted to a single merchant. The agent can queue Chipotle orders on a schedule until the budget is gone, then stop. It cannot buy anything else. Not software subscriptions. Not cloud compute. Just burritos.

That is Subramanian's go-to illustration when describing what customers actually purchase from Para. Speaking on the Para agent wallets episode of the On The Margin podcast, she spelled out the architecture behind the joke.

"There's a team called Colossus that lets you provision credit cards backed with stablecoins with super granular permissions on them," Subramanian said. "I could create a stablecoin-backed card and give it $200 a week and just have it buy Chipotle. So it's only allowed to buy my Chipotle bowl every day."

Beneath the fast-food framing is a structural argument that separates Para's positioning from every other company in the agent payment space. The spending policy lives inside the wallet. That distinction, Subramanian contends, is one that every major protocol launched in the past year has failed to address.

The last time we had a new financial rail was probably credit cards in the 70s. And so it's probably the most exciting time in many of our, if not most of our careers to be building in either FinTech or crypto.

Five Protocols, One Missing Piece

The past nine months have produced a remarkable pile-up of agent payment standards. Stripe and Tempo released MPP. Coinbase launched its x402 protocol as an internet-native payment layer for machine-to-machine transactions. Visa published its Trusted Agent Protocol. Google introduced AP2. Stripe-backed Circuit & Chisel created ATXP. Every major player in payments has staked out territory.

Subramanian's critique is not that these efforts are misguided. It is that they all answer the same question while ignoring a different one. Every protocol in that list addresses how a transaction gets routed across rails. None of them, she argues, address who gets to authorize the transaction before it leaves the wallet.

"The race to own the rails has very much outpaced the actual volume flowing through these rails," Subramanian said. "Like everyone has launched a standard, right. Stripe has MPP, Visa has TAP, there's x402 as well. There's all of this kind of movement towards let's build the rails for agents. But where is the agent volume?"

That volume gap is not a new observation. It matches earlier industry reporting from February 2026, which reached the same conclusion: the standards have multiplied faster than usage. The question she is betting on is whether the wallet layer can close that gap by making authorization programmable in a way protocols cannot.

Why ATXP's Creator Agrees With Her

Louis Amira built the Agent Transaction Protocol from inside the protocol race. He is a Stripe alum and co-founder of Circuit & Chisel. His pitch for ATXP is explicitly about neutral routing across all the competing standards, which puts him in an odd position of agreeing with Para's broader diagnosis even while competing in the space Subramanian is critiquing.

"It is a fascinating web of lots of different protocols and methods and all sorts of acronyms that you and I shouldn't have to memorize or know, and our agents shouldn't care too much about either," Amira said on the same podcast.

His design ambition for ATXP was to build something above the rail wars. "Our aspiration was to scaffold a protocol that allowed all of the different agents to use all of the different payment methods across all of the different sources."

That framing actually strengthens Subramanian's point. If even the people building protocols think agents should be rail-agnostic, then the competitive moat cannot live at the protocol layer. It has to live somewhere else. Para's answer is the wallet.

Our aspiration was to scaffold a protocol that allowed all of the different agents to use all of the different payment methods across all of the different sources.

What Agents Will Actually Buy (and What They Won't)

Commodity purchases. That is Subramanian's blunt answer to where agent-led spending has a real ceiling and where it does not.

Laundry detergent, toothbrushes, recurring grocery orders, travel logistics where your preferences are already logged. These are fine. The agent can optimize on price, timing, and logistics without needing to know anything about you as a person.

Watches, handbags, clothing. Those are different. "Agents are fundamentally about outsourcing a purchase, and anyone who has ever outsourced a purchase knows that this comes with trade-offs," she said. "Maybe all the way from asking your mom to buy you some clothes when she's at the mall, to asking someone to book travel for you. People do have preferences."

Her read is that LLMs will reshape discovery and curation in personal categories. They will surface the right jacket, shortlist the right hotel. But conversion, the moment money actually changes hands, will remain human-controlled in any category where the purchase says something about who you are. The agent economy is real. It just has narrower jurisdiction than the most optimistic forecasts suggest.

How Does Para's Wallet Architecture Actually Work?

Para is a developer SDK, not a consumer app. Companies integrate it to provide their users with non-custodial wallets without building the cryptographic plumbing themselves. Pricing runs on usage: per user, per wallet, or per signed transaction, depending on which integration mode the partner uses.

The agent payment protocols that Para supports include x402, MPP, and several others. The company's REST API, released earlier in 2026, lets regulated financial platforms add blockchain wallet functionality to existing products without replacing their core infrastructure.

The cryptographic foundation is multi-party computation. MPC distributes signing authority across multiple key shares, each held separately, each required to approve a transaction. No single party, including Para itself, can freeze or redirect funds. For agent payments, that architecture becomes the substrate for the permission layer Subramanian keeps describing.

"Wallets are ultimately the authorization and control flow layer of anything that's happening on chain," she said. "So it really is on the wallet layer to also provide that access control of what types of transactions can this agent perform, what kind of transactions can be signed, spending limits, protocol restrictions, chain restrictions, things like that."

The naive design for agent wallets is to give each task its own wallet with a small balance, so a compromise is contained. Para's preference is the opposite: keep the balance in a single wallet and permission different agents' access to it. Less operational overhead, same security posture, better liquidity utilization.

Wallets are ultimately the authorization and control flow layer of anything that's happening on chain. So it really is on the wallet layer to also provide that access control of what types of transactions can this agent perform.

LLMs Are Now Choosing Wallets Without Being Asked

The strangest demand signal Subramanian described is one that most wallet infrastructure companies probably see but few discuss openly. Some of Para's active users never chose Para. Their development tools chose it for them.

"We've had several teams that have now gone live, like we've never spoken to them. But even more so, there have been teams that don't even really know that they didn't really choose a wallet provider to use. This was just what was put in the product by their LLM."

The go-to-market implication is significant. Documentation is the product. If an LLM scaffolding a new project recommends Para because its docs are clear, its API examples are open, and its CLI works cleanly in code, that recommendation happens before any sales call or pricing conversation. Para moved its developer-portal configuration from a web app to a CLI specifically because it makes the integration surface legible to LLMs, not just humans.

That logic, though, does not apply to Para's own security infrastructure. The company draws a clear internal line between code that is meant to evolve quickly and code that is meant to last. "Our product is fundamentally a security product," Subramanian said. "There are certain elements of our own internal code bases where we don't use LLMs." Wallet signing logic and MPC key management fall into the category where human review remains non-negotiable.

Where Stablecoin Rails Are Actually Moving Real Money

Subramanian's observation about which industries are moving fastest on stablecoin adoption cuts against the conventional narrative. You might expect crypto-native fintechs to lead. They are adopting the technology, but they are not producing the most transformative use cases.

Humanitarian aid is. Para works with Qualipay, which processes humanitarian disbursements using stablecoin rails. The previous system involved multiple wire transfers, cash handouts at physical endpoints, and significant reconciliation overhead. Stablecoin settlement compresses that entire chain.

"There's a company called Qualipay that we work with that does humanitarian aid. And humanitarian aid used to consist of many different wire transfers, cash handouts at endpoints, and really just this kind of super messy system. This is where we're seeing there's the most opportunity for something like a stablecoin-based payment rail to save time, save money, improve reliability," Subramanian said.

The adoption pattern she describes is bimodal. Agentic payments in x402 are drawing a lot of builders and experiments, most of them early-stage. The existing fund flows moving through stablecoin infrastructure, by contrast, belong to older, more established financial pipelines. The loudest part of the story gets the least volume. The quiet part is handling real money today.

Call it the two-speed adoption problem. Protocol launches generate headlines. Humanitarian aid disbursements do not. But in terms of dollars cleared and operational impact, the second category is already running laps around the first. Subramanian seems genuinely more interested in the quiet part.

I think people might have guessed that the super tech forward fintechs would be kind of the first to fully move over to stablecoins. But I'm actually finding some of the more transformative use cases are companies that operate completely outside tech in terms of their core user base.