AI-Wiki

Will Agents actually be able to buy things on the internet? - John Collison, Co-Founder of Stripe

7 min read

Collison — Will agents actually be able to buy things on the internet? (Browserbase Navigate, Oct 2026)

▶ Watch on YouTube · Browserbase · 28:33

Will Agents actually be able to buy things on the internet? Stripe Co-Founder John Collison thinks so.

We hosted him as our keynote speaker at Navigate 2026 to talk about agentic commerce, computer use, the singularity, and their recent acquisitions.

— channel description, Browserbase

TL;DR

A ~29-minute keynote conversation on the Browserbase channel, published 5 October 2026, recorded at Browserbase’s Navigate 2026 conference in San Francisco. John Collison, co-founder of Stripe, is interviewed on stage by Molly O’Shea. Browserbase sells headless browsers for AI agents, and Collison says it is “well on its way to becoming a huge Stripe company”: host and guest each have a commercial interest in the thesis. The captions are auto-generated; quotes are cleaned for names only. There are no slides: the stills scan found only stage footage.

1. Agents need computer use, “the ultimate backwards compatibility layer” (0:25–2:20). Stripe processed “$1.9 trillion in 2025, up 34%” (the host’s figure). Personal agents such as Instinct and Muse are “real”. “Filling out a web form is not a value-add activity for anyone.” An agent that knows your shipping address and payment details, and can operate a computer, can “renew my driver’s license for me.” “How long would you need to wait for … the DMV to add an MCP? … it doesn’t matter because computer use solves that.”

2. Agentic commerce rewires selling in two orders (2:20–4:45). First order: “people’s agents are going to be doing a huge amount of the commerce for them.” Second order: AI product research “often actually surfaces niche brands or businesses you haven’t heard of … quite a good story for the long tail.” “AI discovery is the new point of aggregation for commerce”, so incumbent aggregators face questions and the top sellers will turn over. His analogy: Facebook and Instagram ad targeting produced a generation of native sellers (Wish, Teespring, Temu, Shein) that out-adopted big-box retailers with the same tools.

3. Cybersecurity gets worse before equilibrium (4:45–6:52). The new models “have really advanced cyber capabilities”; “over the next 5 years, you’re going to see more breaches than you did over the past 5 years” because many organisations “are going to be too slow in how they adopt the latest capabilities.” Startups are better placed, with “smaller, more contained, more homogeneous tech stacks.”

4. “Day one of the singularity” was 1 January 2026 (6:53–9:30). A Stripe in-joke, set after the “holy crap moment” with coding models at the end of 2025. “AI capabilities are helping with the rate of improvement in AI.” The economic signal he cites: “We’re seeing the highest rates of new firm creation that we’ve seen in recent memory on Stripe.” No figure is given.

5. Developer ergonomics give way to AI ergonomics (9:35–11:40). Stripe spent its history on the developer experience of its API. “Relatively overnight, no one actually really integrates the Stripe API by hand anymore … they send their coding agent to do it.” What helps an agent differs: “it’s very easy for them to … integrate the wrong version of the API.” “We are having to kind of totally change what we’re good at internally at Stripe.” For every industry, “everyone’s chess boards just got a little bit of a jiggle.”

6. The acquisitions, read through AI (11:40–15:20). Bridge and Privy bring crypto-native culture. Metronome is usage-based billing, because “because of AI, everyone has inference costs”. OpenRouter bets that “more than one model will be used inside a business”, routed by workload or by the user’s lifetime value. How much of Stripe’s volume is AI: “a lot”, but hard to measure, since “if you’re not using a lot of AI in your business, you’re probably not paying attention.” Build versus buy: buy when the muscle is new and the market moves fast enough that an internal build would be out of date on arrival.

7. Inside Stripe: AI on internal data, built by its users (17:29–20:20). He cites the jagged-frontier: models do well where there is training data and poorly elsewhere; the organisational version is that consumer apps lack business context. Stripe built an internal tool that gives AI access to internal data, “which is quite complex because you need to have all the privacy guardrails … access control guardrails.” With it, “all our sales teams … completely changed their workflows”, and “they are building their own application”; someone on legal built an agent that ingests new regulation. “People have been talking about personalized software for decades … finally the barrier to entry has gotten low enough.”

8. Computer use as the next discrete hop (20:20–21:10). Transformers, then LLMs, then RL; “good computer use … will come to be viewed as one of these next hops.”

9. The unasked question: bot detection when it is all bots (24:05–25:25). “Many sites do bot detection and bot prevention … but like it’s increasingly all bots. And so, what is the new equivalent of … bot detection … where there is the good bots and then there are the bad bots?” Try Instinct, Grok or Meta’s Muse to see “that new modality for how AI products will work”, but “I think we just don’t know” what the new equilibrium will be.

10. Lessons from 17 years (25:30–28:17). “There’s no shortcuts”: no breakout moment, just growth every year; new products start as “a toy”; stick-with-it-ness (Nvidia as example); the long-term, “infinite game” with partners; and Stripe’s first operating principle, “users first”.

Dynamic-capabilities reading

  • contextual/external-triggers — agentic commerce and AI discovery change how every merchant is found; Collison frames it as the board being jiggled for every industry.
  • strategic-renewal/business-model — Stripe’s core competence (developer ergonomics) loses value as coding agents integrate the API, and the acquisitions (usage-based billing, model routing) follow the cost structure of AI products.
  • digital-transforming/redesigning-internal-structures — sales and legal staff build their own AI applications on internal data, which needed new access-control and privacy guardrails first.
  • Ognibeni predicted that agents would take over search-driven e-commerce first; Collison describes the same shift from the payments side.
  • Amiel & Sanders (HubSpot) give the brand-side practice for the “new point of aggregation” Collison describes; see answer-engine-optimization.
  • Cloudflare’s one-year report measures the non-human traffic behind Collison’s bot-detection question, and builds one answer to it (Web Bot Auth). Stripe co-founded the x402 foundation with Cloudflare and Coinbase, per agentic-web.
  • Chesky (Silicon Valley Girl), from the same batch, names the same personal agents (Instinct, Muse) and argues that chat is the wrong interface for commerce.
  • Ramineni (Fireflies), on the same channel two days later, describes routing across seven to ten model providers and a move from seat to usage pricing, the two practices behind Stripe’s OpenRouter and Metronome acquisitions.
  • Klein (Browserbase), the conference’s opening keynote, gives the host’s own case for computer use. It uses the same DMV example and offers a product answer (Web Bot Auth signing) to the bot-detection question Collison raises.

Linked entities and concepts

Debates and supersession

  • Interested parties. Browserbase sells computer-use infrastructure, and Stripe processes the payments of agentic commerce and counts Browserbase as a customer. The two claims carrying the talk (computer use is the next AI wave; agents will do much of the buying) come with no adoption figures.
  • Long tail or new aggregator? Collison expects AI discovery to help well-reviewed niche sellers. The agentic-web page records the opposite worry from DFI and Ognibeni: the agent operator captures the customer relationship. Both can hold if the aggregator changes and the long tail gains reach under it.
  • Computer use vs protocols. Collison treats computer use as the way around waiting for MCP servers. The wiki’s MCP-oriented sources assume the integration layer gets built; this is a bet on which arrives first, not a contradiction.

What was actually ingested

The full 28:33 keynote from YouTube’s auto-generated English captions, fetched with yt-dlp after the transcript panel failed to render. The Gemini stills scan (manifest, 0 stills) returned no visuals, and a ten-frame probe confirmed two people on stage throughout.

Relationships

Supports

Graph View

Backlinks