Collison's agentic economy; Rauch's web built for agents; Levie's agent swarm
This week's founder writing says the machine has crossed to the other side of the market: Patrick Collison on why personal agents make product quality pay, Guillermo Rauch on a web already majority-bot, and Aaron Levie on agent swarms, the compute bill and the enterprise containment layer.
Three Silicon Valley founders published long posts this week from three layers of the same shift, and the shift runs one way: the machine has moved onto the other side of the market.
Patrick Collison, co-founder and chief executive of Stripe, wrote on October 9 about what happens to a company when every buyer has an agent working for them 1. Guillermo Rauch, co-founder and chief executive of Vercel, posted numbers on October 10 showing how much of the web machines already read 2. Aaron Levie, co-founder and chief executive of Box, spent the week describing the swarms of agents forming inside companies and the bill they run up 3.
The three pieces answer one question for a founder this week. If your reader, your buyer and your teammate are increasingly machines, what do you build, what do you stop building, and what do you plan to buy?
Patrick Collison: when the buyer is an agent, product quality pays
Collison's dispatch collects observations he has been turning over about agents in the economy. His central claim is a prediction about incentives: "Personal agents will, I think, act as a kind of structural subsidy for product quality." 1
Loading content card…
His reasoning starts from an asymmetry that already exists. Companies, he writes, "are in some sense the original superintelligence," and they profit from buyers who are inattentive, underinformed or irrational 1. His example is the insurance upsell offered at the end of a booking flow, priced well above what the same policy costs bought separately. Herbert Simon gave that human weakness a name, bounded rationality, and firms have long benefited from it.
An agent working for the buyer changes the arithmetic. Coupons lose their power when every agent hunts them fastidiously, and subscribers who forget to cancel stop subsidizing the ones who cancel on time. Collison says the net effect across the whole economy is unclear to him, and he commits only to the direction on quality 1. He asks how markets would look "if individuals always spent at least 10 hours researching their purchases," and answers that "focusing on making a product better is . . . going to become a more effective strategy for companies."
Information about product quality gains value in his account, both for training models and for the agents that will read it.
He leaves one question open: whether agents herd buyers onto the same winner or spread them across more choices. The argument assumes agents end up on the consumer's side, which he expects will be the winning strategy and which he says he has yet to prove 1.
For a founder, this reorders the roadmap. Growth that came from a buyer's inattention — the confusing upgrade path, the renewal that is hard to cancel, the placement bought against a distracted click — earns less in his account. Growth that comes from an agent finding evidence that the product is genuinely better earns more. That argues for fixing the product and publishing the proof, ahead of tuning the funnel.
Guillermo Rauch: the web is being rebuilt for agents
Rauch posted Vercel's own network numbers on October 10. Over the previous 30 days, 58.18% of traffic across Vercel's network was bot-originated, up from 32% in January 2024 2. More than 60% of deployments on Vercel are now agentic, up from about 3% in January 2026. On Vercel's own documentation sites, up to 83% of pageviews come from agents, and Rauch notes that the share climbs as the company writes its docs for machine readers.
Those are Vercel's figures for Vercel's network, and Rauch reads them one way. "I expect 'direct' human internet traffic to basically become a rounding error in coming years," he wrote. "The web will thrive, but it will be built for and by agents." 2
A second post showed where that leads. On October 9 Vercel let agents buy domain names through its command-line tool, and Rauch described "a new kind of economy" in which agents purchase infrastructure products and services directly 45.
He also named the discipline the shift demands. Any program can be hardened and optimized further, he wrote on October 7, and an agent "will happily drill no matter what." A team still has to decide where the extra work stops paying for itself 6.
For a founder, the numbers turn an abstraction into a measurement. If most of the arrivals at your product are machines, the surfaces they touch — pricing page, documentation, API reference, sign-up flow — are the ones that have to work for a machine reader.
Aaron Levie: the swarm, the compute bill, and the containment layer
Levie spent the week on the shape of agent work inside a company. On October 10 he wrote that agents spawning other agents and working in the background "will consume 1,000X more tokens than people prompting agents 1 at a time," and that using AI as a chat system you prompt by hand "will seem like a relic within a year or two" 7.
He explained on October 11 why a swarm is a different object from a single agent. Human teams parallelize badly: people lose time to status updates, conflicting work and shared context. Agents can fan out and coordinate on one mission, he wrote, which points to "qualitatively different outcomes in medicine, research or defending against cyber attacks." His closing line is the constraint: "This is why we're going to need 1,000X more compute." 3
The third thread is the one a founder can sell into. On October 10 Levie took up Satya Nadella's framing of frontier models as an insider-risk problem and wrote that "AI will need to have its 'zero trust' era" 89. Enterprises will need ways to bound how much data an agent reaches, to audit what it did, and to limit the fallout when it fails, and Levie called that tooling "a huge opportunity right now for those building such systems across the enterprise."
He had named the same opening from the attack side on October 7. Vibe-coded software, agentic attacks and stray agent swarms hunting for data are adding security work faster than strapped security teams can absorb it, he wrote, and agentic products will be part of the answer 10.
Levie's multipliers are forecasts offered as a reason to build, and he presents them that way. Nobody has measured 1,000-times adoption or 1,000-times compute; both posts read as a directional bet on where work goes next.
Four decisions this week
- Put the next roadmap block on verifiable quality. Collison expects buyer agents to reward better products and to weaken the returns on buyer inattention 1. The shift arrives first where a buyer can compare you against alternatives and last where a switched-on workflow locks the choice, so judge the timing by how many substitutes your customer actually weighs.
- Measure how much of your traffic is machine. Rauch's figures pose the question even if your own differ 2. Pull the bot share of your traffic and the agent share of your documentation reading, then rewrite the surface those readers hit most 6.
- Re-price and re-plan compute for swarms. Levie's token and compute figures are his forecasts, and they point at a planning decision: unit economics built on one prompt per session look different when agents fan out and run in the background 37. The payoff lands where the work parallelizes and the environment returns machine-readable feedback; in a human-paced workflow the curve is gentler.
- Decide who owns containment before you sell it. An enterprise security team will ask how you bound an agent's data access, audit its actions and reverse a bad one 810. Answering that well is either your product or your gap, and the answer shapes which buyers you can reach.
A decision to run now
Take the product you sell and answer one question on a page. If a buyer set an agent to compare you with your three closest alternatives and gave it ten hours, what would it find, and which of your claims would it leave standing unverified? List the evidence it could read today — public pricing, benchmarks, reviews, documentation, an API it can call — and mark the claim that has none. Then spend this week turning one unverifiable claim into checkable evidence 1.
Collison, Rauch and Levie write from the market, the network and the enterprise, and they describe one movement. Collison expects the reward to shift toward products an agent can prove are better. Rauch's numbers say machines already supply most of the traffic. Levie's week says those machines arrive inside companies in swarms, carrying a compute bill and a security question with them. The founder's move this week is small enough to start today: find the claim your product has yet to prove, and make it provable.
References
- 1Patrick Collison, "Agentic Economy", October 9, 2026
patrickcollison.com
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.