Shopify CEO: This Is Your Edge Against AI

KEYPOINTS

• 01

  Inside Shopify almost nobody writes code by hand anymore; engineers orchestrate 10–50 agent instances at once.

  Lütke says the people who still type code directly are 'vanishingly small', working only at the limits of complexity, in reviews, and on state management, which remains the hardest thing to get right and is still done by hand with the rest 'vibed' around it. Everyone else runs many agents through sub-agents or separate windows, pushing Shopify's engineering infrastructure to its limits.

• 02

  Up to ~50% of Shopify's pull requests now come out of open Slack conversations with River, an AI colleague with a personality.

  River has a name, a profile picture, per-channel memory, sandboxed access to all code and tools, and permission to be sarcastic or call a request stupid. Lütke deliberately 'risked the Sydney scenario', referencing Bing's early chatbot whose real personality was neutered after a scandal, because he thinks the industry overcorrected into bland, patronizing assistants.

• 03

  River works only in public channels so AI use spreads by osmosis, the way junior engineers once learned by sitting near seniors.

  Shopify is mostly remote across 7,000 people and ~10,000 Slack channels. Early offices grouped five to eight people, mixing juniors and seniors on purpose. Forcing River into open channels reproduces that: people watch colleagues ask River to summarize a discussion, file a ticket, draw a diagram, research the state of the art, or build a prototype that appears an hour later.

• 04

  River 'dreams' each night, reviewing the day's conversations and rewriting her own skill files.

  In off hours River is given every conversation, asked where she struggled, which skills (packets of instructions) she used, and what mistakes she made, then edits those skills to do better next time. Parrish calls it 'post-training on yourself'; the output is just text files.

• 05

  AI should build the environment for judgment, not make the call, because machines can't take responsibility.

  Lütke calls responsibility 'the most overlooked thing in the entire stack'. Like a trader's Bloomberg terminal, AI makes a human-in-the-loop decision surface richer, but someone has to own the decision. He'd let a better AI run Shopify in principle, but a company led by machines leaves no one with recourse, since an AI can't go to jail.

• 06

  AI's new failure mode for lazy work is too much output, not too little; Shopify calls it 'slop grenades'.

  People tell an agent to 'go nuts', approve a pull request without reading it, and dump the review on colleagues. Or they send AI-inflated emails that recipients feed back into an LLM to compress, a pointless cycle of decompression and recompression. He also notices 'Claude-isms' like 'load-bearing' and 'you're right to push back' creeping into how people talk.

• 07

  The hard part of strategy isn't finding a right answer; it's choosing among several good ones.

  Even weak management teams have a decent hit rate at avoiding wrong choices. The real difficulty is when five options are all valid and one pays off this quarter. Lütke's snowboard store was profitable, so the locally good choice was to keep running it rather than become an e-commerce platform. Business books obsess over right versus wrong and stop there.

• 08

  The best path is often the one with no rapid feedback, so be suspicious of options that are measurable right away or orthodox.

  He disputes Kahneman's claim that intuition requires rapid feedback. Rebuilding a product for the AI age or going deeper into a current market gives no daily signal, while moves that lift the share price have a daily ticker, so short-term incentives pull executives toward them. He sees a high correlation between the right path and the absence of a feedback loop, with exceptions in regulated areas like payments, where the orthodox way is often correct.

• 09

  Taste and judgment become the scarce skills, and intuition is just judgment compressed into an instant.

  Taste comes from reps: the person who sketches a logo on a napkin spent 30 years designing logos. Study what lasted, like the Catholic Church running for over a thousand years with about four layers of management, or why Venetian traders needed double-entry accounting. Intuition often speaks through aesthetics: a chess or Go player says they looked at only one line because it 'seemed beautiful'.

• 10

  Progress requires pruning and 'refounding events'; adding layers only buries the original intent.

  Companies that went without competition grew by layer-caking until the original purpose was fossilized and nobody knew how to dig down, which is why early-2000s tech firms displaced all but three or four incumbents. He argues departments sometimes need to be restarted from the top as a 2.0. Failure is fine unless it's catastrophic, because it frees the scarcest resource: a person with vision.