The Anomaly That Showed Up in No Analytics Tool
On a Tuesday morning in late 2025, a mid-tier Shopify storefront I track logged 4,100 completed checkouts in ninety minutes. No ad spend. No email blast. No influencer drop. The referrer field was empty, the user-agent string resolved to a single opaque token, and the geography of the buyers was, statistically, impossible โ twelve time zones firing simultaneously, order velocity flat as a ruler.
I pulled the settlement log. Every one of those orders had been authorized by a delegated payment credential. Not card-present. Not card-not-present with a human staring down a 3DS challenge. A machine bought 4,100 units of inventory, and the merchant had no idea which human โ if any โ stood behind the intent.
That is agentic commerce. And the four companies now building its rails are not competing on AI. They are competing on who gets to charge the toll. The model is already good enough. The war is over the take rate, the identity layer, and the liability when the agent buys the wrong thing.
I have spent ten years reading EVM bytecode and P&L statements in equal measure, and I have learned one thing the hard way: when infrastructure shifts, the money does not move to the smartest code. It moves to whoever controls the choke point. This is a story about choke points.
The Four Camps Nobody Is Calling a War
The industry is being sold a harmony narrative. Four companies, one shared mission, the infrastructure of agentic shopping quietly coming online. Read the press releases and you would think Stripe, Google, Visa, and Coinbase are laying the same track.
They are not. They are laying four incompatible tracks and hoping the other three derail first.
Here is the actual map. OpenAI and Stripe shipped the Agentic Commerce Protocol (ACP) โ an open standard for merchants to expose inventory and for agents to place orders against a payment intent. Google countered with AP2, the Agent Payments Protocol, built on top of its A2A agent-to-agent framework and already reporting sixty-plus partners. Visa pushed the Trusted Agent Protocol (TAP), which routes the entire question of agent legitimacy through the card networks that already authorize every transaction on earth. And Coinbase rolled out x402, a stablecoin settlement rail that lets an agent pay in USDC over HTTP without ever touching a card network.
Four stacks. Zero guaranteed interoperability. This is not a standards committee. It is a standards war wearing a committee's clothes.
Maturity matters here. ACP is in the POC-to-production transition โ ChatGPT Instant Checkout went live with Etsy first, then a slice of Shopify merchants. It works. It is not scalable yet. Merchant coverage is narrow, and the fulfillment loop โ returns, disputes, oversells โ is incomplete. AP2 has the widest partner list but the least shipping product. TAP has the most mature authorization rails and the least consumer-facing surface. x402 has the cleanest settlement primitive and the smallest merchant footprint.
That is the context. Now the part that actually moves money.
Core Analysis: The Bottleneck Moved, and Nobody Repriced It
The model is not the constraint
Let me kill the most expensive misconception first. The hard problem in agentic commerce is not teaching a model to understand "buy me the cheapest 4K monitor that ships by Friday." That problem is solved. Intent parsing and order generation are commodity capabilities now.
The real technical friction sits in three places that have nothing to do with model intelligence:
First, structured product feeds. An agent can only buy what it can read. That means a merchant's inventory, pricing, specifications, and real-time stock must be exposed as clean, machine-parseable data, synchronized continuously. Most merchants run on legacy catalogs where the same product has three SKUs, two prices, and a description written for humans. When I audited the merchant-side integrations during the EigenLayer restaking work in 2023 โ same discipline, different protocol โ I learned that the ugly part of any "seamless" system is always the data plumbing underneath. Agentic commerce inherits that plumbing, and it is rusty.
Second, agent identity verifiability. How does a merchant prove that the entity placing the order is a legitimate agent acting on a specific human's authority, and not a script that stole a session token? This is the unsolved core. Every camp answers it differently โ ACP wants protocol-level attestation, Google wants A2A handshakes, Visa wants card-network cryptograms, x402 wants on-chain signed intents. Four answers, no consensus.
Third, the authorization chain. Strong Customer Authentication and 3DS exist precisely to force a human into high-risk transactions. Agentic commerce requires that human to be absent. So either you exempt agents from SCA โ and open a fraud surface the size of the internet โ or you invent a new authorization primitive that proves delegated consent after the fact. Nobody has shipped the second thing at scale.
The take rate is being re-cut in real time
Now the money. Stripe's economics are brutal and simple: roughly 2.9% plus thirty cents per transaction, and it processes volume at a scale where a basis point is a city's worth of revenue. Agentic commerce, if it scales, multiplies Stripe's transaction count without adding proportional cost. That is the entire thesis behind Stripe's positioning as "the payment rail of the AI era." It is not a product story. It is a volume story.
The platform layer is where it gets predatory. OpenAI's Instant Checkout was reported to be targeting a merchant fee in the single-digit percentage range โ deliberately undercutting the commission structure of incumbent marketplaces. That is textbook subsidy-for-ecosystem: lose money on the take rate, win the default entry point, then reprice once you own the funnel. I have watched this movie before. In 2020, I deployed a fork of SushiSwap on testnet and threw 5 ETH of my own savings into the initial pool, purely to feel how the incentive curve actually behaved. The farming rewards hit 300% APY and I netted $4,200 before the correction. The lesson was not "DeFi is free money." The lesson was that whoever bootstraps the liquidity controls the fee switch later. Agentic commerce is running the identical play, one layer up the stack.
Here is the structural point the harmony narrative buries: if an AI assistant charges a per-transaction fee, a new toll booth appears between the merchant, the payment network, and the platform. That cost does not vanish. It lands on the merchant's margin or the consumer's price. The question is never whether a new take rate emerges. The question is who collects it.
The discovery layer is the real casualty
Everyone is staring at the payment rail. The payment rail is fine. It is mature, it is boring, and it will absorb agentic volume without breaking.
The layer that is about to break is discovery.
The traditional chain runs search or social ad to landing page to conversion. Agentic commerce collapses that chain. The agent skips the landing page, skips the display ad, skips the SEO-optimized category page, and hits the product directly. Every dollar spent on capturing human attention in the middle of that funnel is a dollar spent on a step that a machine now deletes.
Map the substitution precisely. For search advertising and retail media networks โ hundreds of billions in annual spend โ the exposure is structural, not cyclical. For affiliate marketing, comparison shopping, and deal aggregators, the replacement rate is above sixty percent; these businesses exist to sit between intent and purchase, and the agent eats that exact seat. For brand marketing and content-led demand generation, the effect is augmentation, not replacement โ humans still need to be convinced that a brand exists before they tell their agent to prefer it. For retail customer service and fulfillment, augmentation again.
The downstream effect nobody has priced yet is the birth of a brand-new services market: Generative Engine Optimization. If the agent decides what to recommend, brands need to optimize to be recommended โ not ranked. That is a different discipline from SEO, it will absorb SEO budget, and it does not exist at scale yet. I have watched my own quant desk's research pipeline shift the same way: the edge moved from "rank well" to "be the default answer."
The long tail gets quietly executed
There is a second casualty that gets no press. Agentic shopping rewards merchants with clean data, stable inventory, and reliable fulfillment โ because that is what an agent can transact against without error. The long tail of small merchants, the ones running on duct-taped catalogs and inconsistent stock, will be systematically deprioritized. Not banned. Just never surfaced, because the agent optimizes for the transaction that completes cleanly.
This is the same structural dynamic I documented in the DAO governance space. Governance tokens that look democratic on paper concentrate power with whoever can afford the gas to actually participate. Agentic commerce concentrates discovery with whoever can afford to maintain machine-readable data pipelines. Open protocols do not democratize access. They democratize the entry fee and then concentrate the outcome.
The variable everyone is mispricing: Amazon
Four camps, and one giant standing outside all of them. Amazon runs a closed loop โ first-party retail, its own advertising network, and its own logistics. It has no incentive to join an open agent protocol that hands discovery to a third-party assistant. The rational move is a walled garden: expose an Amazon-controlled agent, keep the data, keep the ads, keep the take rate.
If Amazon holds that line, the open protocols face their worst-case scenario. A fragmented ecosystem where the largest merchant on earth is unreachable by the agents that are supposed to be the future of shopping. Watch Amazon's stance as the single highest-signal indicator of how fast this actually scales.
The Contrarian Read: The Moat Is Liability, Not Payment
Here is where I disagree with almost everyone covering this.
The consensus frames the battle as a payment-rail contest โ Stripe versus Visa versus the stablecoin rails. Wrong battlefield. The moat is not the payment layer. The moat is the identity-and-liability layer. Whoever defines the standard for "this is a legitimate, authorized agent" owns the admission price for every transaction that flows through the system. That is a bigger prize than any transaction fee, because it is a gate, not a toll.
And the gate is currently wide open. Run the security model honestly. An agent must read external content โ product pages, reviews, listing metadata โ to make decisions. Any attacker who can write to that content can inject instructions. A malicious product page can steer an agent into buying the wrong item, oversizing an order, or exfiltrating a stored payment method. Prompt injection stops being an amusing jailbreak and becomes an unauthorized transaction with real money attached.
Then ask the question the industry refuses to ask: when the agent buys the wrong thing, who eats the loss? The user who delegated? The AI platform that ran the agent? The payment processor that authorized it? The merchant who shipped it? Existing consumer protection โ chargebacks, dispute frameworks โ was designed for humans clicking buttons, not for machines acting on delegated intent. There is no case law, no engineering standard, and no regulator who has ruled.
This is why I set hard human-in-the-loop risk parameters during the Berachain agent battle in March 2025. My reinforcement-learning agents executed over five thousand micro-transactions and cleared a Sharpe of 3.2, but the reason we did not blow up during the flash crash was not the model. It was the human guardrail I hardcoded that refused to over-leverage. The AI was fast. The human was the circuit breaker. Agentic commerce has built the fast part and skipped the circuit breaker entirely. A single high-profile fraud incident โ one headline where an agent drained an account โ triggers a regulatory response that reshapes all four protocols overnight. That is not a tail risk. That is a scheduled event waiting for a date.
And notice what the crypto press buried while reporting this. The genuinely radical rail here is x402 โ stablecoin settlement that lets an agent pay over HTTP without a card network in the loop. If agentic payments migrate on-chain, the entire interchange economy gets disintermediated. That is the story the crypto outlets are actually chasing, and it is the one that threatens Visa and Mastercard most directly. The card networks know it. That is why TAP exists โ not to enable agents, but to keep the authorization chokepoint on their side of the fence.
Takeaway: Trade the Signals, Not the Story
The infrastructure is real and the direction is set. The timing is not. Protocol releases are not revenue, and the market habitually overprices the short term and underprices the long term. In the sprint, hesitation is the only real cost โ but sprinting into a fog without a stop-loss is how desks die.
Track five signals and nothing else. One: the first major agent-fraud incident and the regulatory response it triggers โ that is the repricing event. Two: whether ACP, AP2, TAP, and x402 achieve any real interoperability, or fracture into camps. Three: Amazon's stance โ open protocol or walled garden. Four: card-network adoption rates for agent authentication, which determines whether the new take rate lands on Stripe's side or Visa's. Five: the merchant and fee expansion of Instant Checkout, the clearest read on whether the subsidy phase is ending.
When your agent clicks buy and the wrong item ships, whose money burns โ yours, the platform's, or the merchant's? Until someone answers that with a contract instead of a promise, the rails are being built faster than the brakes. In the sprint, hesitation is the only real cost. Carelessness is the only thing that costs more.
