Open-Source AI Needs to Get Serious

Agentic payments are coming. It’s time for open-source agents to step up or their days are numbered.

Written by Matt Corallo

At this point, we can all agree that AI agents are more capable than humans at many basic tasks. We might not want to admit it, but like the fact that your phone knows more about you than your spouse, facts are facts. However, a huge bottleneck to agentic domination remains: agents still can’t reliably buy things. Whether it’s a domain for their lobster-themed religion (a real thing that no Spiral team members subscribe to) or a flight to an upcoming conference, modern internet payment systems are built on technologies that are actively hostile to agents or, indeed, any bots.

This isn’t news. In fact, many companies from Visa to Stripe to Coinbase to Google and OpenAI, are all developing “agentic payments standards” and pushing their adoption. This leaves us at a critical juncture: as large AI labs start to offer hosted agent services, the payment standards they are pushing will not be open to anyone to implement. Whether they use proprietary, hosted models (or not), locally hosted and open-source agents allow for greater exploration and competition in the agent marketplace. If open agents don’t get serious about integrating and driving adoption of open payment rails, they’ll ultimately be boxed out, leaving only a few players allowed to build the agents we all use.

Nearly every payments industry player is trying to position itself to own the platform for agentic payments. Visa is working on an “Intelligent Commerce” product that is as yet undefined but focuses on keeping humans in the payments loop, limiting agents’ ability to act autonomously. OpenAI and Stripe announced the Agentic Commerce Protocol (ACP), an “open standard” that requires permission to implement and lacks major pieces in its specification. Google, on the other hand, announced AP2 and Coinbase announced an extension of it for crypto called x402. AP2 and x402 may be the furthest along in public-agency payment designs, but they only represent an API structure on which payment APIs for specific payment methods can be built. After x402’s announcement, the GitHub repo was flooded with PRs to support the payment APIs for every random token and blockchain you’ve never heard of.

A high-level API, like AP2, that supports every payment scheme, doesn’t actually solve anything, of course, because an LLM doesn’t actually care about specific APIs. Your agent doesn’t care if the API is REST, XML-RPC, or even something built on hand-rolled cryptography. It’s perfectly capable of calling anything that can be sufficiently specified.

What does matter is having funds available for a compatible payment method. A spec that allows for any payment method, when clients will only implement and have funds available for one or two, and merchants will similarly only accept one or two, solves absolutely nothing; payments will always fail due to a lack of payment method overlap. The x402 maintainers appear to have recognized this, and Coinbase is specifically pushing for USDC on the Base blockchain, where Coinbase conveniently collects all the interest on the tokens people are using to pay. Similarly, ACP only allows multiple payment schemes in theory, documenting only sharing credit card numbers with trusted merchants who’ve presumably signed agreements with Stripe and OpenAI.

Meanwhile, Google’s commitment to the crypto-first x402 world appears quite limited. They only tacitly endorsed it as an extension in their AP2 announcement, and haven’t appeared to make any moves towards stablecoins or any other crypto-anything since. And why would they? Google is planning to spend over 175 billion dollars this year, much of it on AI-related initiatives. Given that AI will eventually drive a nontrivial portion of total consumer spend, becoming the payment gateway for that and cutting out Visa to take that sweet 2% of every transaction for themselves would be a great way to see good returns on their investment.

On the flip side, open-source agent developers are still stuck in a “why can’t I just give my agent my credit card number and call it a day?” world. This doesn’t work for a few reasons. First of all, the chargeback mess from people declaring charges fraudulent because their agent shared their credit card number with a scam site, or even because their agent made a purchase they ultimately didn’t want, would bring Visa’s network to its knees. More importantly, the internet’s payment systems are built on a substantial investment in anti-bot technologies, captchas, and tons of effort to ensure bots can’t buy things. Whether an agent is acting on behalf of a human or not, these technologies don’t simply go away, and removing them from existing payment systems would expose merchants to the unacceptable fraud and chargebacks that led to their creation in the first place.

Given that the existing internet needs to upgrade to support agentic payments by offering purchasing paths that aren’t filled with captchas and bot detection anyway, the opportunity is wide open for new agentic payment rails to be built. As Google, OpenAI, and others start adding purchase integrations directly into their chat or hosted agent interfaces, they will be forced to focus on direct merchant integrations with those that offer purchase APIs without bot-hostile limitations. Whether they handle their own payment processing without Visa’s fees or require specific payment integrations and negotiated contracts, the net effect will be that only agents at the large AI labs can buy things.

This leaves open-source and local agents in a lurch. For an agent to be maximally useful, it needs the ability to buy things, but with merchants focused on building integrations with the proprietary payment methods pushed by the large AI labs, agents running other software will be left hobbled. Even when a few large labs control the best models, payments that drive vendor lock-in ensure there can’t be material competition (or creativity) in new agent form factors and use cases.

Given the headline-grabbing announcements about Stripe’s APIs now having options for USDC-on-Base x402, the devs not still stuck on credit cards appear to believe that we’ll live in a world where Coinbase owns both the platform that much of future commerce will happen on (Base) as well as all the revenue from interest on the float for them (in the form of USDC interest). This is rather naive—Stripe has APIs to accept many forms of payment, and merchants haven’t yet materially moved to enable accepting any specific one, let alone turn off much of their existing anti-fraud and anti-bot technologies to allow agents to use it. More importantly, Stripe has also partnered with OpenAI for ACP, offering proprietary agentic payments outside of x402.

It’s time for a serious attempt at open payments for open agents. Whatever you think of bitcoin, it was, in fact, made for this. Payments that are not built around third-party gatekeepers, unsustainable chargeback systems, and inevitable KYC roadblocks are exactly what open agents need. Stablecoins are great, but many businesses, especially internationally, simply aren’t going to accept the future of commerce being built on a platform where one company gets all the interest revenue and a single private key can arbitrarily seize funds.

While nascent, there are already a decent number of global merchants that accept bitcoin, whether for domains or VPS services that might enable agents to be self-hosted, or even for flights and gift cards at major retailers. Of course, some of those remain agent-incompatible due to widespread captchas and the internet’s general hostility towards bots, but some work fine. Better yet, there are many payment processors and wallets that hold dollars but make and accept payments in bitcoin, converting between dollars and bitcoin when needed to avoid any exposure to bitcoin’s price volatility (one of the main arguments for stablecoins to begin with).

Unlike nearly every other payment method, the fact that bitcoin wallets can be created without jumping through KYC hoops that explicitly prevent bots, be funded from many different sources (or even receive funding through an agent charging for services it offers), and work without requiring a human-negotiated and signed contract with a company or without fully trusting a single company (or single private key) to stay online and not seize your funds makes bitcoin the only compelling option for open AI agents to build on. Some have already started the work, with Moneydevkit and the L402 protocol leading the way on agentic wallet integrations, but there’s a lot more work to be done.

Agents should be reaching out to merchants with anti-bot websites and asking them to support agentic commerce (with payment methods that don’t expose them to fraud or chargebacks) instead of simply failing and waiting for everyone to build support for proprietary agentic commerce payment rails. Open agent tooling needs to push open payment methods by default, rather than ceding the ground to large AI labs that will ultimately box them out. And, most of all, we all need to use the only open payment method available, free of trusted third parties or vested interests.