Tobi Lütke: The Skills AI Makes More Valuable

Shopify CEO Tobi Lütke describes agents as shared participants in software work, not merely private coding tools. His main boundary is responsibility: agents can gather evidence, generate changes, and expose more options, but a person still has to choose, review, and own the result.

Evidence basis: This summary uses the saved English auto-generated captions, captured September 19, 2026, without an original-audio check. Shopify practices, adoption figures, and outcomes are Lütke's account, not independent measurements. Product details and forecasts reflect the recording date.

Make agent use visible to the team

• Shopify's River agent works in shared Slack channels. Lütke says River can use company tools and code, answer questions, and propose pull requests from team conversations. He reports that up to roughly half of Shopify pull requests now begin in shared-chat conversations rather than traditional individual implementation. That is a speaker-reported estimate, not a measured productivity result. 07:15–08:56(opens in a new tab)
• Colleagues learn by watching agent use. Lütke says Shopify restricted River to open channels so colleagues could see how others prompt it, ask it to research, summarize discussions, create tickets, or prototype ideas. His intended mechanism is workplace learning by observation, replacing some of the informal exposure lost in distributed work. 08:56–10:53(opens in a new tab)
• River reflects on mistakes to improve its instructions. Lütke says Shopify gives River its conversations during off-hours and asks how its skills—packets of instructions—could be improved after mistakes. He describes the output as skill and instruction text files, but does not explain how changes are checked or put into use. 10:53–12:48(opens in a new tab)

Use agents to prepare judgment, not replace it

Separately, Lütke uses a personal AI chief of staff to prepare decisions. It assigns other agents different perspectives, such as data analysis, research, business, and engineering, then combines their findings into an audio briefing. The aim is to inform his judgment, not hand over the choice: he argues that machines cannot take responsibility for the result. The interview provides no comparison showing that this process produces better decisions. 12:48–15:43(opens in a new tab)

That responsibility also applies to everyday handoffs. Lütke warns that AI lets someone produce more without doing the work of checking or explaining it. He calls unreviewed pull requests and inflated messages “slop grenades”: the sender saves effort by making colleagues check the code or extract the point from a long message. 15:43–17:40(opens in a new tab)

Cultivate taste through depth and repeated practice

Lütke expects taste and judgment to become more valuable as generation becomes easier.

• Develop taste through practice and close study. His example is a designer whose quick logo sketch draws on decades of work. He also recommends studying why designs and institutions endure: what human problems they solve and how their systems address people's limitations. 30:55–35:11(opens in a new tab)
• Use judgment when no option is obviously best. He describes judgment as understanding a system well enough to choose among plausible paths and their trade-offs. Agents can help build that understanding; choosing remains a separate task. 35:11–36:15(opens in a new tab)

Consider subtraction, not only addition

Lütke argues that repeated additions can bury the original purpose of a product or department. By “refounding,” he means starting a new version based on what works, rather than retaining every accumulated layer. His recommendation is to consider removing or rebuilding parts of a system, not only adding more. 58:50–1:01:54(opens in a new tab)

Practical implications

These are editorial implications from Lütke's account, not a validated Shopify playbook:

• Keep consequential agent work visible to the relevant colleagues so they can learn from it and question it.
• Review generated changes before requesting review; distill generated messages before sending them.
• Name the person responsible for a decision and its consequences.