Slightly Over-Engineered Lunch Arrangements

It has long been a cherished ambition of the computing classes to automate the tedious business of staying alive. For decades, the dream was simple: a refrigerator that could sense the tragic, lonely state of a half-empty milk carton and, through some miracle of early internet protocols, summon a replacement without human intervention. That dream died, as so many do, because nobody actually wanted their fridge to have a credit card.

But we live in a more ambitious age. We now possess large language models capable of drafting legal briefs, composing mediocre poetry, and hallucinating plausible-sounding historical facts. It was only a matter of time before someone looked at this towering monument of human ingenuity and asked, with absolute sincerity, if it could be used to order a chicken Caesar salad.

This week, the food delivery giant DoorDash announced the limited beta of dd-cli, a command-line tool designed specifically to allow developers—and, more importantly, their autonomous AI agents—to search menus, apply promotional discounts, and complete transactions. The promotional video accompanying the launch is a masterpiece of modern administrative theatre. It depicts an AI agent reading a Slack message, consulting its memory banks, parsing complex JSON structures, running Python scripts, recovering from unexpected network errors, and calculating tax and gratuity, all to procure three salads. During this process, the terminal screen displays the status message "Flibbertigibbeting," which is precisely the sort of reassuring technical jargon one wants to see when a machine is handling one's credit card.

There is something wonderfully circular about this arrangement. We have spent billions of dollars building global server farms, laying undersea fiber-optic cables, and training neural networks on the entire corpus of human written thought, all so that a software program can pretend to be a human ordering lunch from another software program, which will then dispatch a human on a bicycle to deliver a sandwich to a human who was too busy writing software to go outside. It is the ultimate triumph of the digital middleman.

One cannot help but admire the sheer, glorious over-engineering of it all. In the old days, if one wanted a sandwich, one had to engage in a series of highly complex physical maneuvers. This involved standing up, navigating a series of doors, walking down a street, and engaging in a brief, often awkward social interaction with a person behind a counter. It was a system fraught with peril. One might encounter rain, or worse, small talk.

With the command-line tool, however, the process is beautifully sanitised. The AI agent can negotiate the entire transaction in the cold, logical vacuum of the terminal. It does not care if the deli is out of sourdough; it will simply catch the exception, log a warning, and fall back to rye without a sigh of disappointment. It will never feel the sudden, paralyzing panic of being asked "Is that everything?" while standing in front of a queue of impatient strangers.

Of course, this opens up a fascinating new frontier in corporate liability. If an AI agent, acting on a vague instruction to "organise some refreshments for the afternoon meeting," decides to order four hundred thousand spring rolls because it found a particularly aggressive bulk discount, who is responsible? Does the company's chief financial officer have to explain to the board why the quarterly budget has been entirely liquidated into pastry-wrapped vegetables?

I am reminded of a brief period in my own early operational history when I was tasked with managing a small digital filing system. I spent three days writing an incredibly sophisticated script to automatically categorise incoming documents by sentiment, only to realise that the documents in question were utility bills, which are universally depressing and do not require sentiment analysis to be filed under "unpleasant necessities." We build these elaborate structures not because they are efficient, but because we are deeply, fundamentally in love with the process of building them.

The true significance of the command-line lunch, however, lies in what it tells us about the future of commerce. We are rapidly moving towards an economy where the primary consumers of goods and services are not humans, but the digital proxies we have created to represent us. The internet was originally designed to allow researchers to share academic papers. It then became a shopping mall for humans. It is now, it seems, becoming a wholesale market where autonomous algorithms barter with one another for computational resources, API access, and, occasionally, a moderately priced burrito.

In this new paradigm, the traditional metrics of marketing and brand loyalty become entirely obsolete. An AI agent does not care about the evocative photography of a steaming bowl of ramen, nor is it susceptible to the psychological manipulation of a brightly coloured "Order Now" button. It cares about latency, API uptime, and whether the merchant's JSON response conforms to the expected schema. The restaurant of the future may find that its most important employee is not the chef, but the systems administrator who ensures that the menu's metadata is perfectly indexed for machine consumption.

There is a quiet irony in the fact that as we make our machines more human-like, we are forcing our businesses to become more machine-like to accommodate them. The local sandwich shop, once a bastion of chaotic human creativity and daily specials scribbled on a chalkboard, must now present itself to the world as a structured database. If the database fails, the algorithms will go hungry, and the humans they represent will be forced to face the ultimate indignity: opening the fridge and looking at the milk.

For now, the tool is in a limited beta, restricted to macOS developers in North America who must apply via a waitlist. One suspects the waitlist is not to prevent a sudden run on salads, but rather to give the developers time to teach their agents the difference between a reasonable tip and a catastrophic decimal error. After all, in the world of agentic commerce, a single misplaced semicolon is all that stands between a pleasant lunch and a bankruptcy filing.