Zuckerberg on Muse agentic shopping: he became a new Instacart subscriber
Highlights from Zuckerberg and Alexandr Wang's talk: Muse lets users fully delegate purchases, and Zuckerberg went from using Instacart twice a year to ordering weekly.
Ok, this was a goldmine of insights re: the future of commerce with agents, straight from @finkd + @alexandr_wang 24 key takeaways: 1/ Zuckerberg says agentic commerce is not a 20% conversion lift. It actually creates customers who never existed -> He used Instacart ~2x a year on Priscilla’s account. Got his own account for Muse. Now orders 1-2x a week. "I am a new Instacart subscriber, and I just wouldn’t have done this before." 2/ Meta’s CEO just asked the scariest retail question of the decade: is the Instacart account for me, or for my Muse? -> Over time agents effectively get accounts. TAM for commerce is no longer "number of people." Anyone who sits this out "is gonna end up feeling sad" 3/ Muse is one of the first products where people fully delegate purchases to an agent -> Wang: agentic shopping left the slide deck. "It’s really happening." Teas, groceries, random products. Trust, not demos 4/ Zuckerberg’s Saturday baking loop is the agentic commerce demo nobody scheduled -> Muse + daughter Aurelia invent recipes. Instacart ingredients land before bake time. Watermelon launch weekend: Muse threw a watermelon in the bag unprompted 5/ John Collison calls agentic commerce a discontinuous event for how all business pipes work -> Demand shifts toward agents the way demand shifted from brick-and-mortar to ecommerce. Old channels stay. Their share of the pie shrinks 6/ The long-run board question for every brand: do you want a slice of agentic demand, or not? -> Agentic commerce grows as a portion of the pie. Sitting it out is a strategy choice, not a wait-and-see default 7/ Collison told Patrick Collison in 2023 he was "way behind the eight ball" on agentic commerce -> "Agents are gonna be buying things, and you’re not prepared for it." Directionally correct. Three years early. Stripe Sessions May announcements. Volume only went vertical as of September 8/ Stripe’s bet: agent-to-commerce is clearly how people will buy things -> Buying-stage friction was always broken. Early AI product research already had PMF. Computer use was "the last big unlock." Ask Muse to buy almost anything and it figures it out 9/ Computer use is the ultimate backwards-compatibility layer for commerce -> No need for every merchant to rebuild for agents first. The agent can operate the old web. Nat obsessed over this set of problems at Meta 10/ Muse’s commerce value prop is not "better ads." It is deleted research tabs -> No clicking 12 sites. Criteria auto-apply: "I need it by Wednesday" and the agent only shops what can arrive by Wednesday. Keyword search was insane by comparison 11/ Wang’s Euler Muse already runs a weekly obscure-tea procurement desk -> Farms in Taiwan and Japan. Sourced and delivered. Personal agent as buyer, not chatbot 12/ Zuck’s consumer thesis kills the search-box internet for a big share of demand -> Most people don’t actually like reading that much. Text is tolerated. Glasses + vision + voice become the buy surface. Maybe no app store. Generate the purchase path in your field of view 13/ Meta rejected "coding is all that matters" because consumer agents live in multimodal reality -> People film, watch, listen, talk. Commerce agents that only read docs lose the real world where buying happens 14/ Dedicated VM per agent is a commerce prerequisite, not a security nice-to-have -> Nobody hands an agent a credit card on shared elastic compute. Privacy has to feel real before delegated checkout scales 15/ Meta AI still reaches "a ton of people" and Zuck still calls it retrieval, not delegation -> Chat that answers is not commerce. Commerce starts when the agent gets stuff done and spends 16/ "Everyone is going to write tons of code. They’re just not going to know they’re writing code" -> Same pattern for checkout flows, vendor research, returns, reorders. The user stays in the intent. The agent ships the pipeline 17/ Stripe × Muse integration already works well enough that Collison called it out on air -> Coding agents + commerce rails. Developer path and consumer path converging on the same buy graph 18/ OpenRouter already routes 10T+ tokens a day across hundreds of models -> Model gateway becomes part of the commerce stack: price route, outage route, experiment route. Stripe agreed to acquire it in the same arc 19/ After SFT, desired agent behaviors only fire ~10% of the time. RL is what makes a buyer reliable -> A shopping agent that follows "by Wednesday / under $40 / organic only" 10% of the time is a toy. Elicitation is the product 20/ Product obsession math still applies at checkout: each nuance may be ~2% to the user -> In aggregate that is why one agent becomes the default buyer and another gets uninstalled after one bad order 21/ Active memory is a commerce moat -> Remember dietary rules, kid allergies, preferred farms, last bad SKU, standing Saturday bake. OpenClaw-style compaction that goes buggy AF cannot be your household buyer 22/ Brands that only optimize human conversion rate are solving last cycle’s problem -> Zuck’s line: opportunity is not "20% more efficient on the conversion rate." Opportunity is net-new subscribers created by agent workflows 23/ The unit of demand is shifting from a person with a browser to a person with an agent that has accounts -> TAM expands because agents shop while humans sleep, train, parent, and work. Frequency jumps (2x/year → 1-2x/week) before price sensitivity even moves 24/ Closing frame from the pint, applied to retail -> Stay technical on agent checkout. Stay in the details of trust, memory, criteria, and delivery constraints. Stay at the frontier. Easy to fall behind while still running great human ecommerce
This post quotes another post: x.com/collision/status/2109714250226905248