Over a Cheeky Pint, the Agentic Commerce Playbook

Over a Cheeky Pint, the Agentic Commerce Playbook

Mark Zuckerberg and Alexandr Wang spent two hours and fourteen minutes at a pub talking about Muse Wonder, and the part that matters most isn’t the model — it’s what the model buys. Zuckerberg’s thesis, distilled from the 24-takeaway thread Seijin Jung mined from the conversation: agentic commerce is not a twenty-percent conversion lift. It creates customers who never existed.

The proof is embarrassingly personal. Zuckerberg used his wife’s Instacart account roughly twice a year. He got his own account for Muse. Now he orders one to two times a week. “I am a new Instacart subscriber, and I just wouldn’t have done this before,” he said. The agent didn’t make Instacart’s funnel better — it manufactured a net-new subscriber out of someone who barely used the service. Multiply that pattern across every category of retail and the total addressable market stops being “number of people” and becomes “number of people with agents that have accounts.” Agents shop while humans sleep, train, parent, and work.

John Collison, hosting, called it a discontinuous event for how all business pipes work — demand shifts toward agents the way it once shifted from brick-and-mortar to e-commerce. Old channels don’t die; their share of the pie shrinks. The board-level question for every brand becomes binary: do you want a slice of agentic demand, or not?

The underlying machinery is what makes the thesis credible rather than just vibes. Four load-bearing pieces emerged from the pub:

1. Computer use is the backwards-compatibility layer. Every merchant doesn’t need to rebuild for agents first, because the agent can operate the old web. Wang framed computer use as “the last big unlock” — ask Muse to buy almost anything and it figures out the existing checkout flow. This is the same reason Stripe’s bet is agent-to-commerce: the buying-stage friction was always broken, and early AI product research already had product-market fit.

2. Elicitation is the product, not the model. After supervised fine-tuning, desired agent behaviours only fire about ten percent of the time. It’s reinforcement learning that makes a buyer reliable — a shopping agent that follows “by Wednesday, under forty dollars, organic only” one time in ten is a toy. Relatedly, product obsession still applies at checkout: each nuance may only be two percent to the user, but in aggregate that’s why one agent becomes the default household buyer and another gets uninstalled after a single bad order.

3. Active memory is the moat. Remember the dietary rules, the kid’s allergies, the preferred farms in Taiwan and Japan (Wang’s Euler Muse reportedly runs a weekly obscure-tea procurement desk), the last bad SKU, the standing Saturday bake. Zuck’s Saturday baking loop — Muse and his daughter inventing recipes, ingredients landing before bake time, a watermelon thrown into the bag unprompted — is the demo nobody scheduled.

4. Trust infrastructure comes before scale. Nobody hands an agent a credit card on shared elastic compute: a dedicated VM per agent is a commerce prerequisite, not a security nice-to-have. And the gateway layer matters — OpenRouter already routes more than ten trillion tokens a day across hundreds of models, which makes model routing part of the commerce stack itself (price route, outage route, experiment route).

• The unit of demand is shifting from a person with a browser to a person with an agent that has accounts. Brands optimising only human conversion rate are solving last cycle’s problem.

• Frequency jumps (two times a year to one-to-two times a week) arrive before price sensitivity moves — the opportunity is net-new subscribers created by agent workflows.

• Computer use lets agents transact on today’s web without waiting for merchants to rebuild; the agent layer is the compatibility shim for all of commerce.

• The durable moats in agentic commerce are active memory, elicited reliability (RL, not just SFT), per-agent private compute, and model routing — not a better ad.

Seijin Jung’s full 24-takeaway thread · Wang: annual performance review, live over a pint

Ramp’s “Glass”: Learn Things Once

Last week, David Senra asked Eric Glyman a simple question: “How does Ramp move so fast?” Today came the answer — a product called Glass, and a four-word thesis behind it: “We learn things once.” The launch letter is really an essay about why giving every employee AI is not the same as becoming an AI-native company.

The opening parable is brutal: “It’s like giving every employee a Steinway and expecting a symphony. What you actually get is a lot of expensive noise.” One week in May, Ramp burned $1.5 million in tokens. The CFO didn’t like telling Glyman, and Glyman told the world anyway. The diagnosis: companies lose what they learn as fast as they learn it. Somewhere in the company, someone figured out the perfect sales-deck prep; someone else wrote the prompt that turns a messy customer call into a clean account plan; an FP&A leader built a workflow that catches every budget variance. All of it lives in one person’s chat history and dies when they close a tab, go on holiday, or leave.

Glass is the fix: instead of starting with an empty chat, you start with everything your company ever figured out. Glyman’s demo story is the kind of compound-interest vignette every operator dreams of: he asked Glass how many UK customers Ramp added since launch, and it knew to exclude US companies with UK offices — because Finance taught it that months ago. Nobody explained a thing. Then the numbers: 323 employees have made 6,500+ edits to what Glass knows. Feature requests used to be copied by hand into Salesforce, about 300 a month; in May an account manager saved a prompt to Glass that pulls them straight from customer calls — now they capture 20,000 a month, another prompt prioritises them, a third emails customers when something they asked for ships. More than 5,000 of those emails have gone out.

Two mechanisms here deserve a second look, because they’re the parts most companies skip. First, the Ramp Router: every request runs through a model router that picks the right model for the job, open or closed — “when a better one ships next week, you just use it.” When Ramp moved its own work onto it, their bill fell 30% overnight. Second, the taxi meter: every project has a meter, so you see what work costs while it’s happening, not when the invoice arrives, and you can set limits before they’re crossed. That $1.5M week in May is the cautionary tale that sells the meter.

• The AI-native gap isn’t model access, it’s institutional memory: companies lose what they learn as fast as they learn it because knowledge lives in chat histories.

• The compounding version is when every employee’s work edits the shared brain — Ramp’s feature-request pipeline went 300 to 20,000 a month from one saved prompt.

• Cost control is a product feature: a visible meter plus a model router cut Ramp’s own bill 30% overnight. Unmetered AI spend is how you get a $1.5M week.

• Model-agnosticism as strategy: “when a better one ships next week, you just use it” — the router absorbs model churn so the company doesn’t have to.

Glyman’s launch letter (X Article) · ramp.com/blog/introducing-glass

Muse Wonder, the Sequel: “We Didn’t Train on 3D Demo Games”

The Muse Wonder discourse entered its second week with a disarming admission. Responding to criticism that Wonder is being discounted because it wasn’t reinforcement-learned in 3D environments the way some rivals were, Alexandr Wang conceded the point outright: the team, in his words, simply didn’t explicitly train the model on 3D demo games — and yes, the whole team now feels a bit silly about it. Honesty as a launch strategy: own the missing eval, don’t relitigate the benchmark.

The second beat is the one with teeth. Quoting Zuckerberg’s “Muse Wonder” post, Kavin wrote: “Anyone with enough compute and clear intention can catch up fast.” It’s the clearest statement of the 2026 frontier dynamic yet: the moat isn’t the weight file, it’s the clarity of the target plus the compute to chase it. Editions 15 through 19 of this paper have been circling the same thesis from different angles — the model war is over, the war is for the surface, the harness, the memory. Wang’s admission is the coda: even the labs are learning which evaluations matter only after shipping.

Wang’s reply · Kavin on compute + intention

Be Profitable. Run Maximal Agentic Loops.

Garry Tan’s operating advice for this cycle fits in one line: “Be profitable and run your company with maximal agentic loops.” He was quote-posting Rohit Mittal’s thread on why Series A and B are the hardest rounds for founders today — the era when founders launch vertical AI SaaS companies into a market that reprices them every quarter. Tan’s prescription reframes the fundraising conversation entirely: profitability plus agentic loops is the fundraising strategy, because a company that compounds its own labour doesn’t need to sell its cap table to grow.

The mechanism is worth spelling out. An agentic loop is any workflow where an agent does the work, the output is verified, and the learning feeds back into the next cycle — Glass’s feature-request pipeline above is the canonical example. “Maximal” means auditing every repeated human workflow and asking whether it’s loop-shaped yet. Profitable plus looped means the company’s unit economics improve with scale instead of degrading with headcount — the exact inverse of the default startup trajectory, and the reason Series A/B investors are suddenly the hardest to convince.

And the hiring corollary arrived the same morning. Quote-posting a16z’s “optimism wins” image, Arnav Kumar wrote that the chronically negative are easy to spot in hiring and avoid — but the real damage to startups comes from somewhere subtler (his post cuts off mid-thought, which is almost the point: the damage is the quiet kind). Read it as the human complement to Tan’s machine advice: agentic loops compound execution, but only an optimism-filtered team points the loops at something worth compounding.

Tan’s post · Arnav on the optimism filter · a16z: “optimism wins”

Meesho: Making E-Commerce Everyone’s Business

Vidit Aatrey sat down with McKinsey senior partner Sathya Prathipati to tell the story of how two 24-year-olds built one of India’s leading e-commerce marketplaces — roughly 250 million annual transacting users, a 2025 IPO that raised $603 million with shares jumping sixty percent on listing. The through-line of the interview is a question of where you look: in 2015, Bangalore tech circles thought e-commerce was done. In Aatrey’s hometown of Meerut and co-founder Sanjeev Barnwal’s Hazaribagh, nobody was trading online. So they started with a simple mission: bring everyone in India online.

The early-days material is the best founder-craft in this edition. Aatrey and Barnwal were IIT Delhi roommates; in 2015 Barnwal called from Tokyo, bored of corporate life, and Aatrey said “let’s build something together” — Barnwal resigned the next day. Knowing most startups fail and that innovation is never linear, they made a pact: take the option of shutting down off the table for the first two years. “We were 24-year-olds who had never run a company, failing every day. But that taught us patience.” And on co-founder conflict — the startup killer nobody prices in: “I see founders without that deep trust blaming each other when chaos creeps in. For us, trust was the survival mechanism.”

The product strategy was equally contrarian. While everyone else forced app downloads, Meesho built for WhatsApp users. While the industry charged commissions, Meesho went zero-commission as a permanent mission choice — “we want to make sure that everyone is able to come online and sell anything” — monetising through advertising instead. Logistics got its own wedge with Valmo, bringing new people into the delivery ecosystem. The asset-light pure-marketplace model is what made the IPO timing work: a decade of foundation-laying, then the first horizontal Indian e-commerce platform to go public, ahead of Amazon and Flipkart’s Indian entities.

• The opportunity was geographic, not technological: 2015’s consensus was that e-commerce was “done” because everyone looked at Bangalore. Meesho looked at Meerut.

• Distribution innovation (WhatsApp-first) and business-model innovation (zero commission, ads monetisation) mattered more than any single feature.

• The two-year “no shutdown” pact is an underrated founder instrument — it converts early chaos from an existential threat into a learning budget. Trust between founders was the survival mechanism.

• Ten years of foundation before going public: the IPO was a lagging indicator, not the achievement.

Aatrey’s post · McKinsey: “Meesho: Making e-commerce everyone’s business”

Welcome HomeView: Apple’s Smart-Home Hub Arrives

Two words from Apple’s hardware chief, and a decade of rumours ended: “Welcome HomeView.” At the October 13 “Welcome home” event in Tribeca, Apple unveiled its first smart-home hub — a 7-inch Liquid Retina display in two form factors, tabletop and wall-mounted, with Siri AI as the centre of the experience. The device Apple watchers have called “HomePad” for years finally has a name, a price, and a ship date.

The details matter more than the price. The tabletop model pairs an aluminium body and stainless-steel hinge with a domed speaker base — full-range drivers in the display, a down-firing woofer in the base — in silver, starlight, rose pink and space grey. The wall-mounted version comes in space grey and silver with purpose-built mounts. The display adapts to distance: simpler information from across the room, more detail as you approach, brightness sliding from 500 nits down to 0.3 nits at night. It recognises individual users for personalised apps and services, handles FaceTime calls, music and video, controls lights, locks and thermostats, and shows feeds from up to nine connected cameras.

The strategic read: this is the first Apple product designed around the new Siri AI rather than retrofitted with it — broader knowledge, more natural conversation, personal context drawn from Messages and Calendar when paired with an iPhone, with on-device processing and Private Cloud Compute as the privacy story. No dedicated App Store at launch; third-party widgets can be projected from a nearby iPhone. It ships in the US, Australia, Canada, India, New Zealand and the UK on November 13 (Singapore pre-orders open October 15) — and notably not in the EU. Voice is the primary input; the screen is the confirmation layer. That’s the whole bet: the kitchen counter as the next computing surface.

Ternus’s post

Character.AI Files for Chapter 11; Disney Circles at $20M

The companion-AI era just got its cautionary tale. Character Technologies, the company behind Character.AI, filed a voluntary Chapter 11 petition in Delaware on October 12 (case 26-11603), listing $50–100 million in assets and liabilities — and lined up Disney as the stalking-horse bidder at $20 million for certain assets plus a perpetual IP licence. If no better bid emerges by late November or early December, the platform winds down.

The numbers tell the story of a particular kind of AI-company death. Founded by Noam Shazeer — co-author of the Transformer paper — and Daniel De Freitas, Character.AI had genuine product love: 1.7 million installs in its first launch week, a $1B+ valuation on a $150M raise, 20 million monthly users by October 2024. Then the acqui-hire era: Google paid $2.7 billion in August 2024 to license the models and take the founders back. The company’s own statement blames “significant legal, regulatory, and operational headwinds” — and the creditor list is revealing: Google LLC tops it at $4.1M, followed by law firm Hogan Lovells at $2.1M, DigitalOcean at $1.95M, Fireworks.ai at $814K, plus lawsuits from the Commonwealths of Pennsylvania and Kentucky. When your biggest creditors are your cloud bill, your lawyers, and two state governments, the business model had already lost.

Disney’s angle is the interesting part. The $20M stalking-horse bid isn’t for the company — it’s a baseline offer for certain assets plus a non-exclusive perpetual IP licence, with transition services from the former Character.AI employees Disney hired last month (including ex-CEO Karandeep Anand, now Disney’s CTO). Disney gets the tech; Character.AI gets a court-supervised auction that might still save the platform if a better bid shows up. The platform stays up and billing continues meanwhile.

• Product love isn’t a business model: 20M MAU and a beloved product still died when legal exposure, safety costs and regulatory risk grew faster than the ability to fund them.

• The 2024–26 acqui-hire wave (Google’s $2.7B, now Disney’s people-first $20M) shows where the value actually sat: the team and the IP, not the consumer product.

• Watch the creditor list, not the valuation: cloud bills + law firms + state lawsuits is the balance sheet of a company whose unit economics never worked.

choblin29’s post

199 Indian startups raised $10M+ in 2026

OfflineOnAir’s monster listicle: 199 Indian startups have raised ten million dollars or more this year, headlined by CRED’s $900M Series H (Meta-led, verified — $4.5B post-money, June 2026). The funding winter is officially a selective thaw: late-stage deals dominate, early-stage still gets done. Bookmark the thread; it’s a year’s worth of “who’s real” in one scroll. The full list

Anyone with enough compute can catch up fast

Kavin’s gloss on the Muse Wonder moment: “Anyone with enough compute and clear intention can catch up fast.” The frontier is a function of clarity plus FLOPs, not incumbency. Pair with Wang’s 3D-games admission above. Kavin’s post

The 38-minute airport run

Deedy’s love letter to San Francisco logistics: there is no faster city for flying — house door to airplane seat in thirty-eight minutes flat (9:10pm leave, 9:32pm SFO, 9:40pm gate, 9:43pm boarding, 9:48pm seated). A reminder that great infrastructure is invisible until you time it. Deedy’s post

Annual performance review, live

Wang, on the Cheeky Pint video: “at some point in this video i get my annual performance review live.” Somewhere in those two hours and fourteen minutes, Zuckerberg reviews his newest hire on camera. The internet’s homework: find the timestamp. Wang’s post