Who Wins Agentic Commerce? A Deep Analysis of the Open vs. Closed Protocol War

This fight is really about who controls the relationship between humans and the things they buy.

Last week, two pieces dropped that frame this debate perfectly. Sam Ragsdale published “Open Agentic Commerce and the End of Ads” on a16z crypto, arguing that AI agents will kill advertising and open protocols will win. Hours later, Antonio Garcia Martinez, the man who built Facebook’s first open ad exchange and watched it get killed, responded: open doesn’t always win, most commerce is emotional, and whoever controls demand dictates terms.

We went through 100+ sources across protocol docs, consumer research, payment infrastructure, and market data. The answer to “who wins?” is more interesting than either side admits.

The four protocols at war

Here’s what’s actually out there:

ACP (Agentic Commerce Protocol) is OpenAI and Stripe’s baby. Five REST endpoints for agents to complete checkout with Stripe’s Shared Payment Tokens. Launched September 2025. OpenAI charges merchants 4% per transaction.

UCP (Universal Commerce Protocol) is Google and Shopify’s answer, with 30+ backers including Visa, Mastercard, and Walmart. Covers the full commerce lifecycle, not just checkout. Merchants self-host capability profiles at /.well-known/ucp. Announced January 2026 at NRF.

x402 is Coinbase’s HTTP-native payment protocol using the dormant HTTP 402 status code. Stateless, per-request stablecoin payments. Open source, permissionless. Live since May 2025.

MPP (Machine Payments Protocol) is Stripe and Tempo’s enterprise-grade protocol with “sessions” (think OAuth for money) that aggregate micropayments. Launched March 18, 2026 alongside the $5B-valued Tempo blockchain.

Four different bets on how AI agents should interact with commerce.

Ragsdale’s argument: the end of ads

Sam Ragsdale, who built x402 at Coinbase before founding Merit Systems with a16z crypto backing, makes three structural arguments for open protocols:

1. AI agents can’t be distracted. The advertising business model depends on capturing human attention. Agents don’t browse, don’t get sidetracked, don’t click banner ads. Stack Overflow views are down 75% since GPT-4. Tech news traffic is down 60%. When the “user” is software, the attention economy collapses.

2. Stablecoins solve the 28-year-old micropayment problem. HTTP 402 “Payment Required” was created in 1997 but never implemented because credit card minimums (~$0.30) exceeded the value of micropayments. Stablecoins on L2 chains (sub-$0.0001 on Base) make sub-cent transactions viable for the first time.

3. Open protocols enable edge innovation. “An agent that can only buy from pre-approved merchants is an employee with a corporate card restricted to three vendors. An agent with open protocols is an entrepreneur with a bank account.” ChatGPT checkout requires months of BD, legal docs, and merchant vetting — the AOL model.

Hard to argue with any of that. But it has blind spots.

AGM’s counter: open doesn’t always win

Antonio Garcia Martinez (Director of Base growth at Coinbase, author of Chaos Monkeys, former Facebook PM who shipped FBX) brings the historical pattern-matching that Ragsdale’s piece lacks:

1. Facebook’s walled garden beat Google’s open ads. Google built an open programmatic ad ecosystem with composable design. Facebook opted for a walled garden. “Arguably, Facebook won in the end as Google’s open Web ads are now relatively less valuable and remnant, while the Instagram feed is premium real estate.” AGM has a personal stake in this observation: he bet his career on the open path at Facebook, and it cost him.

2. Consumers don’t want efficiency — they want to waste time. “The big-ticket items in e-commerce are the conspicuous consumption items of lifestyle and fashion, ones where the shopping is much of the consumer experience.” Brands still pay fortunes for SoHo storefronts because the physical retail experience is part of the product.

3. If OpenAI maintains monopsony, openness is irrelevant. If one platform controls the AI consumer experience, it can dictate terms to merchants — the same way Google kept AdWords closed despite championing openness everywhere else. “The only ‘openness’ they care about is having every source of product supply standardize themselves into a product feed that OpenAI can harness.”

4. ACP + MPP is the Apple App Store; UCP + x402 is HTTP. The payments standard that wins depends on “who wins the tug of war to control the consumer experience itself.”

What the data actually shows

Consumer readiness is segmented

The framing of “will consumers let agents buy for them?” is too binary. It’s a spectrum:

• 47% would delegate boring, repetitive purchases (groceries, household supplies)
• 30% would let an AI complete a purchase
• Only 2% would give AI full autonomy for travel
• 95% report at least one concern about AI-assisted purchasing
• 79% say accuracy is what matters most, far above speed (36%)

The generational divide is significant: 75% of Gen Z and Millennials are interested in AI shopping agents, but 62% of Baby Boomers are not interested at all.

Most purchasing is emotional

Harvard research shows 95% of purchasing decisions are subconscious. Gallup finds 70% are based on emotional factors. The brain processes emotional stimuli 3,000x faster than rational thought. People buy on emotion and justify with logic.

This matters because AI agents optimize for rational criteria (price, speed, availability) while most purchasing runs on status, identity, FOMO, and the dopamine hit of anticipation. The engineering mindset that assumes efficiency is the only goal misreads how human commerce actually works.

Manhattan flagships are booming

While pundits predict the death of physical retail, Manhattan retail vacancy dropped to 13.7% in Q4 2025, the lowest since 2017. SoHo and Upper Madison Avenue are below 10%. LVMH, Kering, Prada, and Rolex made 14 major NYC real estate purchases in 2023-2024. Meta just opened a 5-floor flagship on Fifth Avenue.

The physical store has evolved from transaction point to brand temple. As commodity purchases get automated, the experiential premium of physical retail goes up, not down.

OpenAI already retreated

Maybe the most telling data point: OpenAI discontinued Instant Checkout in March 2026 after only ~30 Shopify merchants adopted it. The company couldn’t solve real-time inventory sync, never built a sales tax system, and found merchant onboarding “arduous.”

The pivot to merchant apps inside ChatGPT, redirecting users to retailer websites, is less “agentic commerce” and more “AI-enhanced referral traffic.” If the most powerful AI company in the world can’t make walled-garden checkout work, maybe the debate about who controls checkout is premature.

The payment layer split

Here’s where it gets technically interesting. Agentic commerce payment infrastructure is splitting along economic lines, and both sides are right within their segment.

Traditional card rails remain superior for consumer-facing commerce. Visa and Mastercard are building agent-specific protocols (Visa Trusted Agent Protocol, Mastercard Agent Pay) that work without blockchain. Europe’s first live AI-agent bank payment ran through Santander using Mastercard Agent Pay. Chargebacks, refunds, and consumer trust favor incumbents.

Stablecoin rails are structurally necessary for machine-to-machine micropayments. When an agent makes six API calls at $0.01 each, card network minimums (~$0.30) make the transactions economically impossible. USDC on Base settles for sub-$0.0001. The math doesn’t work on traditional rails.

Stripe’s move: support everything. They co-built ACP with OpenAI, co-built MPP with Tempo, support x402, and are endorsed by UCP. They’re abstracting above the protocol layer, capturing value regardless of which standard wins.

The “open” illusion

Every protocol in this space claims to be open. None of them really are:

• x402: Coinbase’s facilitator handles 70% of all transactions. Single point of failure.
• MPP: Requires Stripe account approval. Not permissionless.
• UCP: “Decentralized” but biggest surfaces are Google Search and Gemini.
• ACP: Apache 2.0 open source, but Stripe is the exclusive payment processor.

The word “open” is doing marketing work in this debate, not technical work. True permissionless commerce, where any developer with a wallet can accept payments with zero gatekeeping, is still aspirational even in the crypto-native stack.

The Farcaster irony

AGM cites Farcaster’s failure as evidence that open social protocols lose. Here’s the irony: Farcaster’s co-founders (Dan Romero and Varun Srinivasan) joined Tempo in February 2026, the blockchain powering MPP, the very protocol in this debate.

The open-protocol builders aren’t abandoning their thesis. They’re pivoting from social (where network effects favor incumbents) to payments (where card network minimums create a genuine market failure). The economic case for openness is just stronger in micropayments than it ever was in social media.

Who actually wins?

Neither. The market bifurcates:

Commodity commerce (groceries, supplies, replenishment): AI agents add clear value. Amazon Rufus is already driving $12B in incremental sales with 60% higher conversion. This segment will be protocol-agnostic — merchants will support ACP, UCP, and whatever else captures traffic.

Experience commerce (luxury, fashion, travel, gifts): Humans retain control. The shopping experience IS the product. Physical retail strengthens. AI assists backstage (inventory, personalization) but doesn’t replace the emotional moment of purchase.

Machine-to-machine payments (agent-to-agent API calls, compute, data): Stablecoins win by default. The economics demand it. This is where x402 and MPP will find real volume — not in consumer-facing pizza ordering.

Stripe captures value across all three segments. The Switzerland of agentic commerce.

What to watch

The best analogy for agentic commerce isn’t AOL vs. HTTP or Facebook vs. Google. It’s email (open) vs. Slack (closed). Different protocols for different contexts, coexisting with interoperability at the edges.

Signals that matter:

• x402 daily volume is currently ~$28K/day with 50% artificial. When it crosses $1M/day of real volume, the crypto-micropayment thesis is validated.
• Dual-protocol merchant adoption: ACP+UCP merchants see 40% more agentic traffic. The winners implement both.
• OpenAI’s next commerce move: the Amazon partnership ($50B investment) and ChatGPT Apps model will define whether OpenAI becomes an aggregator or a referral channel.

The engineering mindset that assumes efficiency is the only goal is bad consumer intuition. The crypto-skeptic mindset that dismisses stablecoin micropayments ignores basic arithmetic about card minimums.

The real disruption isn’t agents buying things for you. It’s agents making the entire supply chain more efficient while humans keep the purchasing decisions that carry meaning, identity, and pleasure.

Both Ragsdale and AGM are right. Just about different parts of the market.