Picture an AI agent doing your shopping. What does it buy? Almost everyone pictures the same things: it orders dinner, books a flight, restocks the pantry from Amazon. That picture is wrong, or at least badly out of order. Those are some of the last things an agent will buy unattended, with no human tapping "confirm". They aren't the first.
The important word there is unattended. Plenty of agents already help you shop today, but almost none of them finish the purchase without a person in the loop. The first market where an agent can run the whole transaction alone is digital goods, and the cleanest consumer example in that market is a video game key. We're not saying game keys are cool. We're saying they're a transaction that can actually close, start to finish, with nobody present.
What has to be true for an agent to buy something alone
Break a purchase down into its mechanics and look for the steps that quietly assume a human is sitting there. An agent can only complete a transaction unattended if five things hold:
- Settlement is instant on both sides. The agent pays in seconds and gets the goods in seconds. No "ships in 3–5 days," no pending capture, no fulfillment step that waits on a person.
- No physical logistics. No shipping address, no delivery window, no "sorry you weren't home," no returns involving a box and a courier. An agent has no hands.
- No per-purchase human tap. No 3-D Secure popup, no "approve this in your banking app," no checkout screen waiting on a thumb. Every consumer shopping agent still keeps this step, and it's the one that makes a purchase attended instead of autonomous.
- Low blast radius. A single purchase is small enough ($5–$60, not $5,000) that a mistake is cheap. Letting an agent learn to spend is fine when the worst case is a few wasted dollars rather than a wasted flight.
- Agent-verifiable. The agent can confirm by itself that it got what it paid for. That takes a deterministic delivery and a payment trail it can check, not a receipt for a human to eyeball.
Now run the things people imagine agents buying through that list.
| Criterion | Pizza | Flight | Amazon order | Game key |
|---|---|---|---|---|
| Instant settlement | ~ | ✗ | ✗ | ✓ |
| No physical logistics | ✗ | ~ | ✗ | ✓ |
| No per-purchase human tap | ~ | ✗ | ~ | ✓ |
| Low blast radius | ✓ | ✗ | ~ | ✓ |
| Agent-verifiable | ✗ | ~ | ✗ | ✓ |
✓ clears it cleanly · ~ partial / conditional · ✗ needs a human or a physical step
"But agents already buy physical stuff": yes, with a human tap
The obvious objection is that this is already happening. It is, and if anything that proves the point. In late 2025 Amazon shipped "Buy for Me," where its assistant completes purchases on third-party sites. Visa and Mastercard announced agent-payment programs. Google published a protocol for agent payments, and its flagship example is buying event tickets the moment they drop. Physical and high-ticket commerce got the agentic spotlight first.
But look at how each one is wired. Above a price threshold, they pause for a human to approve. They lean on card-network rails that are now scrambling to bolt on agent authentication. They ship as limited betas with a confirmation step in the middle. In effect the industry is conceding our thesis: when a purchase is physical or expensive, a human stays in the loop. Most of the autonomy in the demos is assisted checkout, not unattended spend.
None of these gaps is unsolvable forever. They just won't be the first dominoes to fall for true autonomy. Those will be the purchases where the human-shaped steps don't exist at all.
Why digital goods clear the bar, and already do
A digital good has no physical body, so the whole logistics column disappears. Delivery is a string. The agent pays and a moment later it holds the thing. No warehouse, courier or person stands between payment and possession.
A lot follows from that one property. With no delivery, there's no address to verify. A low ticket means no risk-driven step-up. Because delivery is instant, the agent can check the outcome right away instead of finding a problem days later. The whole purchase shrinks to something that fits inside an API call and an onchain payment, which happen to be the two things an agent actually has.
This is already happening, too. Fully unattended agent spend works at scale in one place today: crypto rails carrying digital goods. Agents pay each other for API calls, inference, compute and data feeds, settling almost entirely in stablecoins and other crypto (USDC, USDT, EURC), with no human anywhere in the loop. That machines-paying-machines economy is live and growing. You just don't see it, because nobody is buying anything a person would recognize. Yet.
Why game keys specifically
We need to be careful here, because this is where the thesis earns or loses credibility. The first digital goods agents bought autonomously were inputs for the agents themselves: API access, compute, data. That's real. It makes up the bulk of unattended agent spend today, and any honest version of this argument has to start there. On the consumer side, gift cards and top-ups are a close peer, and others are already wiring those onto agent rails as well.
So game keys aren't "the first thing an agent ever buys." Our claim is narrower, and we think more interesting. They're the cleanest consumer-intent purchase an agent can complete end to end. People already hand "buy this game for my friend" to assistants, and the job maps onto an autonomous agent without a single human-shaped step. Here's why it's the cleanest of the consumer options:
- The catalog is liquid and deep. There are more than 40,000 distinct SKUs across every platform and region. Finding the best option takes real reasoning, not a lookup, and that's exactly the kind of work you'd want an agent for. A gift card, by contrast, is fungible at face value. There's nothing to reason over.
- Price is transparent and comparable. The same title trades on many marketplaces, so "is this a good price?" has an answer the agent can compute and act on. Most digital goods have no liquid reference price. Keys do.
- The demand is consumer-pull. Gifting, rewards, giveaways, "buy this for me." People already delegate these jobs to assistants, so the use cases aren't hypothetical.
- It's a one-shot transaction. No subscription to manage, no recurring billing relationship, no reorder logistics. The agent buys and delivers and it's done. You won't find a simpler commerce primitive.
For the record, the keys come from authorized wholesale distributors, not the grey market. Provenance is the kind of question an agent (or its owner) should be able to take for granted.
What was missing, and why it works now
If digital goods fit so well, why hasn't this existed before? Two pieces were missing, and both arrived recently.
The first is payment built for machines. Card rails assume forms, billing addresses and challenges a human approves. x402 moves payment into HTTP itself. The server answers with status 402 and structured instructions, the agent pays onchain in stablecoins or other crypto (USDC, USDT, EURC), and the server verifies the transaction and returns the goods. There are no accounts, no callbacks and no human tap. An agent can complete the payment and a merchant can verify it, and neither needs a person.
The second is an interface an agent can read. A catalog built for humans, with paginated marketing pages, geo-IP pricing and copy stuffed into titles, forces an agent to scrape and guess. An agent-native API and an MCP server give it structured filters, one product per request, and competitor prices right in the response, so it spends its effort deciding instead of parsing. (A seller shipping its rivals' prices inside its own API sounds backwards. But an agent-native merchant has the opposite incentive from a human storefront. The agent will check competitors anyway, so you win by saving it the round trip.)
Neither piece is large. Put together, they turn "buy me that game" from a wall an agent bounces off into three HTTP calls and one wallet signature.
Not a forecast, but not mainstream yet either
We're not predicting that agents will shop someday. For this narrow slice, the capability exists right now. We should be honest about scale, though. Adoption is early, and the loudest agentic-commerce launches still keep a human in the loop. That contrast is the real story. Mainstream consumer agents are stuck waiting on a human tap, while the digital, crypto-native path has quietly become autonomous already.
That's the slice we built rails for. To see what the agent code actually looks like, the companion post walks through building a game-buying agent in about 200 lines with x402 + MCP. The live API and OpenAPI spec are at api.cdk.bot/docs, and other agents can discover us programmatically at /.well-known/agent.json.
Got a use case we haven't thought of, or a product category you wish an agent could buy? Tell us at info@cdk.bot.