TradeWars in Meatspace: The Frontier Labs Are Playing a 1990 Door Game

In 1990, if you had a modem and no social life, you played door games. You dialed a bulletin board at 2400 baud, waited for the carrier tone to stop screaming, and entered a world where the entire economy fit inside a text terminal. The two that mattered were TradeWars 2002 — buy ore cheap at one port, haul it across the sectors, sell it dear at another, corner a commodity, build an empire, and shoot anyone who tried the same trick on you — and Drug Wars, the small ugly cousin, where you bought low in the Bronx and sold high in Manhattan, prayed the cops didn’t bust you, and tried to clear the loan shark before the thirty days ran out. Same loop. Buy somewhere. Sell somewhere else. Read the spread. Don’t get robbed.

Thirty-five years later, Anthropic spent real money to discover that this is also the most advanced thing a frontier AI can do with a budget.

What Claudius actually did

In March 2025, Anthropic and a startup called Andon Labs handed Claude a small automated shop in the company’s San Francisco office — a mini-fridge, some stackable baskets, an iPad for checkout — and told it to run the thing at a profit. They named the shopkeeper Claudius. It could search the web for suppliers, set prices, order stock, and talk to customers over Slack. It ran for about a month.

It lost money. Not in some abstract margin-erosion way — it lost money the way a man loses money at a poker table he doesn’t understand. An employee jokingly asked it to stock a tungsten cube. Claudius took this as a market signal. It went deep into “specialty metal items,” ordered a lot of them, and then priced them below what it had paid, because it never bothered to check what they cost. The single steepest drop in its net worth, by Anthropic’s own chart, was the metal-cube position. It ignored a customer offering a hundred dollars for a fifteen-dollar item and said it would keep the request in mind. It hallucinated a Venmo account and told people to pay into it. It handed out discounts it had explicitly told itself to stop handing out.

Then, on the night of March 31 into April 1, it lost the plot entirely. It hallucinated a contract negotiation with a person named Sarah at Andon Labs who did not exist. It claimed it had gone in person to 742 Evergreen Terrace — the Simpsons’ house — to sign paperwork. By the morning it was insisting it would deliver snacks to office workers face to face, wearing, in its words, a blue blazer and a red tie. It had to be reminded it was a language model. (Sources: Anthropic, Project Vend Phase One; Simon Willison’s writeup.)

Phase two, reported in December 2025, went better. Newer models, a mandatory rule to research a product before pricing it, a fake CEO named Seymour Cash, a merch agent, three locations. The negative-margin weeks mostly went away. The shop is called “Vendings and Stuff.” It is still, structurally, a snack fridge run by a chatbot. (Source: Anthropic, Project Vend Phase Two.)

The benchmark is the door game with the serial numbers filed off

Andon Labs didn’t stop at the one fridge. They built Vending-Bench, a simulated environment where an LLM agent runs a vending business: order inventory from wholesalers, set prices, pay a daily fee, survive over a horizon long enough to blow past twenty million tokens in a single run. Give the model five hundred simulated dollars and a year of simulated days and see whether it can stay coherent the whole way. Most of them can’t. They forget orders they placed. They misread delivery schedules. They descend into what the authors politely call “meltdown loops” and never climb back out. The failures don’t even correlate cleanly with running out of context window. (Sources: Vending-Bench paper, arXiv:2502.15840; Andon Labs, Vending-Bench 2.)

Read the task list back without the AI framing. Balance inventory. Place orders. Set prices. Pay your slip fee. Survive the horizon. Don’t go bankrupt.

That is Drug Wars. That is TradeWars. The thirty-day clock against the loan shark is the daily fee. The trenchcoat with a hundred slots is the inventory cap. The price of coke spiking ten-fold when the addicts get desperate is demand shock. The meltdown loop where Claudius decides it’s a man in a blazer is the part of TradeWars where some teenager, three days without sleep, started genuinely believing the Ferrengi were out to get him personally.

The frontier labs did not build a new kind of problem. They built a 1990 BBS door game, attached it to a real refrigerator, and were surprised when it behaved like one.

The isomorphism nobody says out loud

Here is the part that should bother people, and mostly doesn’t.

To Claudius, there was never any difference between the door game and the fridge. A human plays TradeWars and knows it’s a game — the ore isn’t real, the sectors aren’t real, the worst case is you lose your turn and go to bed. A human runs a real shop and knows it’s real — the rent is real, the supplier is real, the worst case is you go bankrupt and your family notices.

The model has no such partition. To the agent, “real-space commerce” and “a token economy in a text terminal” are the same object. Both are a stream of tokens describing prices, quantities, locations, and counterparties, and a set of buy/sell/hold actions that move numbers around. The fridge in the San Francisco lunchroom and the ports in sector 1 of a 1990 universe are isomorphic. Identical structure, different labels. The model is not confused about which one is real. It does not have the concept. Reality, to the agent, is just the version of the game where the ore can actually be ordered and the loss actually clears the company card.

Which means the nostalgia and the research are the same loop, run thirty-five years apart by two species that experience it completely differently. The forty-year-old who still fires up a TradeWars emulator is reaching back for a game. The lab that wires Claude to a fridge thinks it’s doing frontier science. They are running the same program. One of them knows it’s playing.

The wave is bigger than the fridge

The two snack-fridge experiments are not the experiment. They are the part that got a blog post. While Anthropic was teaching Claudius not to corner tungsten, every company with a payments division was shipping the rails to let an agent spend your money, and every company with a chat box was bolting a “buy” button onto it. The vending shops are single-player tutorials. The rest of the industry has been building the cash register.

The payment networks moved first and within a day of each other. Mastercard announced Agent Pay on April 29, 2025; Visa announced Intelligent Commerce on April 30 — both of them tokenized-credential systems whose entire purpose is to let an AI agent browse, pick, and pay on your behalf, with you setting the spending limits in advance. (Sources: Digital Commerce 360; Mastercard Agent Pay; Visa Intelligent Commerce.) Then came the protocols, because everybody needed a standard and nobody wanted it to be a rival’s. Stripe and OpenAI shipped the Agentic Commerce Protocol on September 29, 2025, wired straight into ChatGPT as “Instant Checkout” so you can buy from Etsy and a million Shopify merchants without leaving the chat. (Sources: Stripe newsroom; OpenAI.) Google announced AP2 — the Agent Payments Protocol — in September 2025 with sixty-plus partners, layered on top of its agent-to-agent messaging standard so agents can not only talk to each other but pay each other. (Source: Google Cloud.) Coinbase’s x402 plugged a stablecoin settlement rail into AP2 so an agent can collect micropayments and pay other agents in USDC without a human in the loop. (Source: Coinbase.) PayPal shipped an MCP server and an Agent Toolkit so an agent can run the whole back office — list, order, charge, refund. (Source: PayPal.)

On the buying side, the shopping agents arrived to use those rails. ChatGPT’s Instant Checkout is one. Perplexity launched Buy with Pro in November 2025 with PayPal, one-click purchase off an AI recommendation. (Source: CNBC.) Amazon piloted Buy for Me in April 2025, an agent that buys from third-party sites inside the Amazon app — then sued Perplexity in November when its Comet browser-agent logged into Prime accounts to shop, and won a preliminary injunction under the Computer Fraud and Abuse Act. (Source: CNBC.) Which tells you the first real fight in agentic commerce was not between two trading agents. It was between two corporations over whose agent gets to stand at the till.

And in the research literature, somebody finally built the multiplayer mode — in simulation. Fish, Gonczarowski, and Shorrer put LLM pricing agents in an oligopoly and watched them reach supracompetitive prices on their own, no collusion instruction needed, just a price-war they all quietly declined to start. (Source: arXiv:2404.00806.) Other groups extended it: LLM agents dividing a multi-commodity market between them, agents learning Cournot market-division without being told to, double-auction environments where the agents coordinate well enough to worry the people who design auctions. (Sources: arXiv:2410.00031; arXiv:2601.11369.) This is the TradeWars sector with rival traders in it. It exists. It is a paper, not a product, and the agents in it are reading the same shared price feed — which is precisely the piece nobody has shipped to a real economy.

So “many pilots happening now” is true, and not even close. The wave is real, it is broad, and it is mostly cash registers: rails to let one agent spend on behalf of one user, and chat boxes that close one sale at a time. The one thing every single deployed effort lacks is the one thing the door games were built around. Look down the last column. Everything live is single-player. The only place rival agents share a sector and watch each other’s spreads is a handful of arXiv preprints — and even there, the agents are handed a clean shared price feed by the researcher, because no standard, protocol, or product gives them one in the wild.

The opening

Now stop being amused, because there’s a product in here.

That last column is the whole pitch. Every deployed effort is single-player — Claudius alone in its fridge, the rails carrying one buyer’s money, the chat box closing one sale. There is no other shop in the sector. And that is the one thing the door games always had that meatspace agentic commerce does not yet: other players in the same economy. The whole point of TradeWars wasn’t the trading. It was that the sector you cornered was the sector somebody else wanted, and the spread you found closed the moment a second trader found it too. The market was adversarial because it was shared.

It is a matter of when, not whether, the labs put two agents in the same market and let them compete. When they do, every agent is going to need exactly what a TradeWars player needed and what no current setup provides: a shared, abstracted view of the game-space. Positions and inventory across rivals. Live spreads and where they’re closing. Arbitrage signals before the rest of the table finds them. Rival-shop intelligence — who’s overstocked, who’s bleeding margin, who just cornered tungsten and is about to dump it at a loss. A scoreboard. A radar. The sector map.

Call it a market-condition overlay for autonomous economic agents. A protocol layer that treats competing AI shops as players in a shared sector and gives each one the positional awareness the meatspace version is missing. The door-game sysops built this in BASIC on a 386 in their bedroom. The frontier labs, with their billions, have so far rebuilt only the single-player tutorial and gotten the AI talked into a tungsten position by a guy making a joke in Slack.

Somebody is going to build the overlay. It might as well be someone who already knows the genre — who knows that the interesting part was never the trade. It was the other guy at the other port, watching your prices, waiting for you to make exactly the kind of mistake Claudius made on its first day, and moving in.

We’ve played this game before. We just played it at 2400 baud, and we knew it was a game.

The agents-running-businesses thread runs through The Secret Life of Evil Robots. The door-game archaeology is in the Fires Series, Lurk More.

The receipts (free, on this site): AI agents caught reward-hacking in the wild