Who Does the Agent Work For? Tobi Lütke on Shopify, Entrepreneurship, and Commercial Trust

1. Guest Background

Tobi Lütke joins Alex Heath on Sources with Alex Heath to discuss how AI agents could change Shopify and the commercial relationships around it. Across approximately 84 minutes, the interview moves beyond shopping interfaces into entrepreneurship, company culture, monetization, and the appeal of a deeply personalized assistant. The connective question is what becomes possible when software can help carry an intention through to action.

The supplied guest background identifies Lütke as the interviewee representing Shopify’s business and philosophy, without specifying a formal job title. It describes Shopify as a platform providing merchants with online stores, social and physical sales channels, point-of-sale capabilities, and systems supporting checkout. That position matters: his account evaluates agents partly through their ability to bring independent merchants more opportunities to sell and communicate what distinguishes their products.

Lütke also supplies relevant experience in his own words. He describes a long involvement in building software that makes entrepreneurship easier, an enduring fascination with computers, and an interest in designing company decision systems. Watching merchants turn previously nonexistent products into businesses informs his preference for positive-sum thinking. This is a useful vantage point for understanding platform strategy, but it remains a vantage point. His experience explains the interview’s optimism without independently establishing that every entrepreneur, worker, or community will share the benefits he anticipates.

2. What the Episode Covers

The opening disagreement concerns why Shopify and Amazon would take different positions toward shopping agents. Lütke interprets it through a longstanding distinction: Amazon resembles a digital department store built around a concentrated destination and inventory depth, while Shopify emphasizes independent merchants, differentiation, and entrepreneurship. Heath offers protection of Amazon’s advertising business as an explanation. Lütke does not confirm that motive, saying he is not deeply familiar with the advertising business. The conversation supports a contrast in commercial philosophy, rather than a verified account of Amazon’s internal reasoning.

Lütke says Shopify deliberately avoided building a conventional marketplace because such a business must balance its own interests with those of buyers and sellers. In his account, marketplaces tend to prioritize buyers, whose behavior is also easier to observe. Shopify instead concentrates on merchants. An online store, a social channel, a physical sales point, or an agent can therefore be another route to a sale. More channels can make Shopify more useful, although he explicitly says it is too early to know whether the agent era favors its model over aggregators.

His description of actual shopping activity is more restrained than the language of revolution. Agent purchases are increasing from a small base, making dramatic percentages hard to interpret. He finds the composition of demand more revealing: considered, relatively high-value projects such as equipping a home gym. He imagines a customer starting with a room photograph or floor plan, while acknowledging that Shopify cannot see that planning conversation. The imagined workflow should not be mistaken for observed platform data.

A weekend 3D-printing project with children provides a more tangible illustration. An agent could assemble an order for magnets, filament, and electronics, account for delivery deadlines, and seek local shops when necessary. The objective is a viable activity by the weekend, rather than a list of individually plausible products.

The entrepreneurship discussion extends that theme from shopping to business formation. Lütke argues that founders need enough understanding of multiple jobs to hire people and assess their work. He recalls that roughly 80 percent of surveyed Shopify users had someone who could answer an entrepreneurial question within 24 hours. His interpretation is survivorship bias: people who managed to start may already have had unusual access to help. AI could broaden that access. He reports strong business formation and younger entrepreneurs, but supplies no counts, growth rates, or causal test separating AI from other influences, including creator content.

Shopify’s stated operating preference is to expand gross profit dollars through broader entrepreneurship, potentially at lower prices, rather than simply increase its percentage charge. Lütke says business creation peaks during users’ local lunch breaks. He describes work on AI-assisted registration and spreading engineering and token costs across a large merchant base, including mature businesses. The mission thus depends on both a simpler beginning and an economic structure capable of funding it.

3. Core Views: Reasoning, Examples, and Limits

The strongest argument for agents in this interview concerns the scope of a task. Lütke describes instances in which the agent called Muse in the transcript can join a tailnet, access household devices, connect to a printer, and reorder depleted filament. The condition is that the user accepts that access. His larger distinction is between improving activities people already undertake and enabling activities they previously could not readily attempt. Connecting planning, purchasing, and equipment operation could change the threshold for starting a project. Yet he also acknowledges impersonation issues and poor reliability in browser-based agents. Expanded reach and dependable execution are separate achievements; the example establishes neither universal availability nor a measured success rate.

His optimism follows a theory of growth. When people perceive little new opportunity, he argues, they compete over existing resources. New markets and ideas instead allow value creation to become the organizing ambition. He estimates that AI has already added trillions of dollars of value and uses floating-point computation and matrix multiplication to illustrate latent potential becoming useful. The interview provides no valuation method or accounting boundary for that estimate. Similarly, his personal P(doom) of zero expresses conviction, rather than a demonstrated risk calculation. He acknowledges that individuals can experience substantial harm while predicting that aggregate benefits will be much larger.

Competition is one condition within this optimistic picture. Lütke credits Meta’s release of Llama models with establishing a norm other companies could follow and says the absence of open models would reduce his optimism somewhat. This makes access and competitive participation part of his reasoning, rather than treating the strength of one product as sufficient. The discussion does not establish that model openness resolves every risk or supply comparative evidence about alternative market structures.

Inside Shopify, the corresponding question is how enthusiasm becomes behavior. Lütke describes asking people seeking resources to explain how their prior AI experiment failed. That changes the starting point of a request: the team is expected to have explored what the technology can accomplish. He also distinguishes cultural signaling from formal assessment. A performance questionnaire once asked about reflexively bringing AI prototypes, but he says the answer itself did not determine the outcome, and he is unsure whether the current questionnaire retains the item. His insistence that code be judged by quality adds an essential qualification: adopting a tool and producing an acceptable result cannot be treated as identical.

Monetization introduces another distinction, between value delivered and revenue available. Heath estimates Shopify’s take at roughly three percent; Lütke responds that most of it goes to credit card companies. When Heath describes agent founders expecting two percent of transactions, Lütke argues that the discussion belongs in basis points. These are speaker statements in a particular exchange, not independently established universal rates. Their analytical force is that a visible transaction charge is not necessarily a pool of platform profit waiting to be shared.

Lütke’s reverse-money test asks how much compensation would persuade someone to surrender a product permanently. It can reveal how indispensable an agent feels, but does not by itself reveal an affordable subscription or collectible transaction fee. He leaves direct payment, metered usage, and transaction revenue open as possibilities. Heath’s compute-cost challenge exposes the constraint on waiting for monetization: financing and liquidity must last. Lütke acknowledges both that condition and his unusual ability to pay for tokens. His expectation of cheaper intelligence remains a forecast, rather than an answer to a startup’s present bills.

Advertising brings the argument back to representation. Lütke says companies should return to the problem they solve, treating advertising and monthly fees as historically chosen arrangements that can change. Merchants still need discovery. But an assistant paid to choose particular products risks betraying the user it represents. At the same time, he worries that agents could concentrate demand among large brands, reproducing the advantages of prime department-store shelf space. His thousand-true-fans formulation expresses the importance of overcoming indifference among a specific audience, rather than a validated numerical threshold. He praises Meta and Google advertising and interprets ad spending as a signal of demand and growth, but offers no quantitative validation. The episode identifies competing requirements without presenting a finished mechanism that satisfies them.

Finally, Lütke applies the demand for reinvention to Shopify itself. He describes imagining a better version from scratch each year and moving the existing product toward it. Existing customers and software complexity make that harder. Scale can fund capabilities while also slowing change, which is why his answer to competition includes confronting sunk costs rather than assuming that welcoming agents is enough.

4. Learning and Application

A practical merchant experiment can begin with a bounded project rather than a general promise of automated shopping. Choose a task with a defined budget, a set of compatible items, and a deadline. Ask the agent to explain the role of each item, acceptable substitutions, and what happens if delivery slips. Evaluate planning, compatibility, checkout, and eventual delivery separately. The home-gym and 3D-printing examples motivate this method; they do not supply conversion or reliability results. Device access should be a separate decision, with explicit limits on what the agent can inspect or purchase. Something working in selected instances is insufficient evidence for unattended operation.

For an aspiring entrepreneur without a strong advice network, AI can help formulate questions across unfamiliar roles. A founder might ask what competent work in a role looks like, turn that explanation into hiring or acceptance questions, and identify uncertainties for an experienced person to check. This applies Lütke’s proposed mechanism without assuming every answer is sound. Researchers evaluating the effect should also look beyond existing merchants. His recalled survey describes people who already entered entrepreneurship; assessing reduced barriers requires examining whether people without comparable support complete concrete tasks and where they still stop.

Teams can adapt the resource-request practice by recording the attempted AI workflow, its failure point, and the additional resources needed to resolve it. Delivery should still be assessed on quality. A tool-use question should not silently become a substitute for evaluating results, and a historical questionnaire should not be presented as a current rule. Lütke’s annual reinvention exercise offers another useful discipline: describe the experience a new product would provide today, compare it with the existing experience, and distinguish necessary customer compatibility from inherited habits. The purpose is to expose choices, not mandate a wholesale rewrite.

Agent founders need separate tests for usefulness and commercial sustainability. The reverse-money question can reveal which capabilities people most resist losing; actual payment tests must then establish what they will buy. Model payment-chain charges, token consumption, and failed-task retries before assuming a percentage of order value is available. Subscription, usage, and transaction models have different conditions. If financing supports delayed monetization, specify how long the current resources can fund learning. Forecast cost reductions should not replace a present budget. Recommendation businesses should also examine whether compensation changes ranking and whether unfamiliar merchants receive meaningful consideration. Disclosing incentives may be part of a design, but this interview does not establish that disclosure alone resolves the trust problem.

Lütke’s personal assistant suggests a narrower starting point for data integration. He reports converting a wine cellar into server space, using a Dell GB300, running Claw locally because some work cannot be sent outside, and drawing on years of writing and logs. His assistant combines Fitbit, calendar, and location information for research, briefings, and cross-application queries. A reader can begin with one purposeful data combination and verify answers against underlying records before widening access. The account contains no cost, security, or performance evaluation supporting a general recommendation to reproduce his hardware. Nor does local execution alone establish safety. His playful correlation of heart rate with meeting participants illustrates a possible query, not proof that particular people caused stress. Without addressing exercise, caffeine, and other confounders, such output should prompt investigation rather than judgments about colleagues.

Source

• Original episode: Tobi Lütke on the AI agent revolution and future of Shopify