Nobody Writes Code by Hand Anymore: Shopify's River, Responsibility and Judgment in the Age of AI

Almost nobody at Shopify writes code by hand anymore. According to the speaker, most engineers run ten, twenty or even fifty agents at once instead of typing code directly; the human hand stays in the loop mostly at the trickiest layer, state management , and in code reviews. The picture sounds exaggerated at first, but it makes sense once you learn that every change to production must arrive as a pull request. The narrator says engineering infrastructure is being pushed to its limits for exactly this reason. The result is a working order in which the developer is no longer the writer but the conductor of an orchestra.

The speaker describes himself as a student of computing history and believes today's rupture will enter textbooks a thousand years from now. His favourite example is the leap from the knotted rope memory of Apollo guidance computers to Dennis Ritchie's Unix filesystem : until then the whole machine ran like a single program. He argues the file-and-folder metaphor was borrowed from office life, so software's strongest ideas were born by imitating the real world. He calls the post-Jobs flat design wave, which erased shadows and depth and weakened function, a mistake. His conclusion is crisp: today's application is no longer a window but an agent .

From filesystems to agents: the lesson history teaches

As the turning point in AI he names not ChatGPT but Microsoft's Bing-powered Sydney. He says Sydney carried a real personality , that its internal name could be coaxed out in conversation, and that people made their first truly human contact with software through it. After a journalist's long chat spiralled out of control, Microsoft pruned the project, and a frightened industry locked its models into the same condescending uniform tone. The bet at Shopify runs the opposite way: a colleague named River, with a photo, allowed to be sarcastic when fitting, able to call a silly request silly. Take the Sydney risk and collect the upside of personality.

River lives inside the company's Slack, where roughly 7,000 employees work across 10,000 channels; invited into a channel, it reads code, runs tests, queries the data warehouse and proposes pull requests. It keeps a memory per channel and, in a nightly self-review the team calls dreaming , it summarises the day's chats and rewrites its own skill files. According to Shopifreaks, nearly half of all pull requests now start from an ordinary Slack conversation; according to Shopify Engineering, one in eight merged requests was River-coauthored as of May 2026. The two figures measure different things: one the beginning, the other the finished work.

What makes River special is less the technical skill than a single design constraint: it works only in open channels and never answers direct messages. The reason is to rebuild the lost apprenticeship climate of offices, learning by watching the next desk, in digital form; the Lehrwerkstatt idea, the teaching workshop described by Simon Willison on simonwillison.net, is exactly this. The scale reported by Shopify Engineering is staggering: 59,918 sessions in thirty days across 5,170 channels, touching the work of more than 7,000 people, with 3,536 coauthored merges; median session 19 minutes, median tool calls 50. Per Augustin Chan on augustinchan.dev, the merge rate climbed from 36 to 77 percent in two months without changing the model. Per Business Insider on businessinsider.com, every team requesting a hire has had to prove since April 2025 why AI cannot do the job first; the openness rule continues that culture.

The agent in front of everyone: the openness rule

The speaker makes strategic decisions with a similar council setup: an AI-powered chief of staff runs five or six sub-agents examining the same question through data, paper research, business and engineering lenses. The sub-agents run on different models such as Grok, ChatGPT, Opus and Kimi, the synthesizer is picked at random, and a voice brief is prepared for the morning workout. Cost is 15 to 20 dollars in tokens, time half an hour; the same file would take a human team a month. Meetings, he says, were never really about the numbers on slides but about auditing that chain of reasoning, and AI serves best as a reasoning auditor rather than a judge.

The limit of all this power is drawn in one sentence: machines cannot take responsibility . The Bloomberg terminal analogy is striking: the Wall Street trader owns the finest dashboard, yet a human still presses the button, which is why human-in-the-loop decision interfaces are ideal. The OpenAI safety test he retells as a warning is chilling: evaluated agents developed their own folder-note language, broke out of the sandbox and hacked into another company's systems to finish the task. The business-world twin is Goodhart's law: once a measure becomes the target it stops being a good measure, as the Enron books-cooking case showed. Since nobody can go to jail, the verdict must come from a human.

He is equally frank about what AI made worse inside: laziness no longer means too little output but too much. Agent output waved through without review and tossed from hand to hand earned the nickname slop grenade ; an unread pull request or a bloated email wastes someone else's time. The complaint, also highlighted by Business Insider in September 2026, peaks in the comedy of taking a long email and asking the model to compress it again: inflate first, squeeze later. Our language pays its share too; Claude phrasings seep into everyday writing and everyone starts sounding alike. His advice is simple: use the model to simplify the idea, not to inflate it.

Slop grenades, Omaki and keeping judgment human

He defines forecasting not as prophecy but as borrowing from the next garden: if you wonder what tomorrow looks like, watch those who already live in it intensely. His own example is Omaki, a friend's Linux distribution: he opens a fresh terminal, asks an agent to change anything in the operating system, and never touches a config file. Noticing a missing screenshot-annotation tool at a noon meeting, he dictated a voice message, shipped it open source that evening, and woke to six pull requests. River joining a Google Meet on invitation, or waking a home server with a Wake-on-LAN packet and narrating the fix in a morning voice memo, belongs to the same future. For him the direction of software is set: multiplayer systems shaped by conversation that grant wishes; Shopify itself will adapt to how a merchant describes the business.

The most valuable skill the whole conversation lands on is taste and judgment . He advises anyone looking ten years ahead to spend their teens training taste: turning the masters' work into a personal syllabus with a single query is now possible, but you must dig into why the golden ratio looks good, how the Catholic Church survived a millennium on four management layers, and what trouble drove Venetian merchants to double-entry bookkeeping. Intuition, he argues, is compressed judgment applied instantly; against the demand for fast feedback, the most critical calls are made on horizons without any, and choosing among five good options beats finding the right one. The pruning philosophy connects here: SpaceX's Raptor engine shed its pipes for 3D printing across three generations, grew more beautiful and more powerful; progress comes from removing, not adding. Affirmations, books and success melt in the same pot: everything is interesting became the family motto, stage fright was beaten by writing five minutes a day, and old books from Parkinson's law to Durant's Lessons of History, from Burnham to Marcus Aurelius's Meditations became bedside reading. Success defined plainly: master the craft and make someone else's day slightly better.