Vol. I · Founder Memo
The Agentic Commerce Brief
May 14, 2026
Strategy memo · Two-person team · Mercari alumni

The rails are
real. The company
is the question.

Five business ideas for a Staff Engineer and a Director of Product who codes — ranked by what actually survives contact with Stripe, Google, and the protocol wars now unfolding.

Audience
TwoStaff Fullstack + PM/coder
Domain advantage
MercariListings · C2C · Japan
Core thesis
Agentic CommerceOptional: generative UI
Protocols in scope
x402 · MPPUCP · ACP · AP2 · TAP · A2A
§0 The lede

The agentic commerce thesis is thinner than the marketing suggests.

Stripe Sessions 2026 announced 288 products. Google launched UCP with Shopify in January. Coinbase has x402 live on Base. The narrative is loud. Design around it, not into it.

  • Real consumer GMV through agent surfaces in May 2026 is, generously, a rounding error against e-commerce GMV. ChatGPT shopping has not displaced Amazon. Google AI Mode has not displaced Google Shopping.
  • ACP, UCP, and the dominant rails are owned by the people who own the surfaces — OpenAI, Google — and the merchant base — Shopify. A two-person team building on top of those protocols is a tenant on someone's railroad.
  • x402 and MPP are interesting at the API-economy layer, but that's a different business from merchant commerce. Buyer personas, unit economics, and timelines are unrelated. Don't conflate them.
  • The interesting question isn't "what can ACP do." It's "what does ACP create demand for that ACP itself won't deliver — and that the rails-owners don't want to build themselves?"

Mercari trained you on three things that the rails-owners are structurally weak at: messy long-tail listings where attributes are inconsistent and items are unique; C2C/peer trust and dispute mechanics with no merchant-of-record to absorb risk; and Japanese commerce, which is large, idiosyncratic, and almost entirely unaddressed by the current protocol crop. Every idea below leans on at least one of those three.

The five ideas

Ranked by payoff × tractability.

For your specific two-person configuration. Star the ones you want to revisit.

0 / 5 starred
01
Top recommendation

Yoast for Agent Commerce

Agent-readiness SaaS for the merchants Shopify won't help.

Wedge exists: high You can win it: moderate

The premise: ACP, UCP, and MPP all assume merchants will produce clean, structured, agent-consumable catalogs — with the right metadata, pricing semantics, return-policy fields, and tax treatment. Shopify is doing this for its merchants for free, because Shopify co-authored UCP and partners on ACP. Everyone else is on their own.

That "everyone else" is enormous: WooCommerce (~30%+ of e-commerce sites globally), BigCommerce, Magento, Salesforce Commerce Cloud, custom builds, the long tail of Squarespace and Wix, and — crucially — non-Western platforms (BASE, Stores.jp, Rakuten, Cafe24, Lazada/Shopee sellers). These merchants must become agent-readable or they will be invisible to the next wave of search and shopping traffic.

The product, concretely

  1. Catalog connectors — start with WooCommerce, BigCommerce, manual CSV.
  2. LLM-powered normalizer that takes messy product data ("Mens running shoe new BLACK size 10 ⚡FREE SHIP⚡") and outputs clean, attribute-rich records mapped to UCP/ACP schemas.
  3. Output feeds: UCP-compliant for Google, ACP catalog file for Stripe, schema.org/Product for organic crawlers (ChatGPT, Perplexity).
  4. Agent View Preview — your generative-UI hook. Merchant types a query; sees how their products render in ChatGPT, Google AI Mode, and a generic agent. Uses design.md-style declarative specs + LLM rendering.
  5. Analytics: what agents are actually fetching, skipping, and querying for.

Why you specifically: Mercari's entire job is making messy seller-supplied data legible to humans and machines. You know the failure modes — attribute-stuffed titles, photo-only listings, inconsistent sizing, region-specific categorization. This is two months of MVP for a Staff Fullstack + a coding PM. It's six months for any team that's never touched commerce data.

Monetization
$49–$499/mo SaaS tiered on SKU count. Land at the prosumer ~10k–100k SKU merchants.
Defense vs. Stripe / Shopify
Stripe's free upload helps Stripe merchants. You serve the WooCommerce, Rakuten, multi-protocol middle. Switzerland of agent feeds.
02
High variance

The agent-native C2C marketplace

The actually ambitious one. Pure Mercari DNA.

Wedge: moderate You can win it: low–moderate Asymmetric upside

ACP and UCP are merchant-of-record protocols. They presume a SKU, a brand, an inventory count, a static price, a returns policy underwritten by the merchant. They are very bad at: one-of-a-kind items (vintage, used, collectibles, art), negotiated pricing, peer trust without a merchant-of-record, and unstructured listings (photo + caption rather than full attributes). That's the Mercari problem, and the 2026 protocol stack doesn't address it.

Build a marketplace whose listings are designed from day one to be agent-queryable: embeddings, structured attribute extraction from photos, a normalized C2C dispute layer, and integration as an MCP server / x402 endpoint / ACP-compatible catalog. Agents in ChatGPT, Claude, and Gemini browse and buy from your inventory.

The differentiator vs. Mercari/eBay/Vinted: those are app-native consumer marketplaces. You're headless. Your homepage is an MCP server. Sellers list via mobile app or via ingestion from Mercari/eBay/Yahoo Auctions. Buyers never visit a "site."

Wedge: one vertical, one geography

  1. Japanese used/vintage fashion sold internationally via agents. Mercari already proved Japanese vintage has huge international demand.
  2. Ingest from Mercari and Yahoo Auctions JP. Normalize. Translate. Expose as an MCP server.
  3. Settle in USDC via x402 for international buyers; JPY locally.
  4. Agent layer collapses language and discovery friction. Single query → cross-border purchase.
Monetization
5–10% take-rate + payment markup. Much higher ceiling than SaaS, much harder to bootstrap.
Honest risk
Two-sided market. Brutal. Mitigation: start as a meta-marketplace — ingest public listings from existing platforms before recruiting native sellers.

Honest assessment: the high-variance bet. If you're risk-averse, skip it. If you want to swing, nothing in the deck leverages your background more.

03
The gen-UI play

Generative-UI merchandising SDK

Adaptive product rendering for agent surfaces.

Wedge: moderate Monetization: low–moderate

This is the most direct play on the "core: agentic commerce, optional: generative UI" prompt. When an agent shows a product to a consumer today, the rendering is generic. Merchants have no brand presence, no merchandising control, no ability to highlight a feature that matters to the specific query. ChatGPT shows a Glossier card; Glossier had zero input into what was emphasized.

Build a JS SDK + server that merchants embed. When an agent surface requests a product, your SDK returns not just data but a generative UI spec — a declarative description (think design.md) the agent's renderer uses to display a query-appropriate, brand-coherent product experience.

The magic moment

  1. Query: "is this shoe good for marathon training?" → SDK returns a card whose primary visual is the cushioning cross-section, with stack height + drop + plate stiffness called out, plus a 30-day return badge.
  2. Query: "casual everyday running shoe" → same product, different render. Primary visual is the colorway, secondary is comfort score and reviewer demographics.
  3. One SDK install. Adaptive rendering. A/B-tested. Brand-coherent.

Why you specifically: a Director of Product who codes is the right founder profile for design-y SDK work. The Staff Fullstack ships the spec compiler and renderer.

Why I rank it third: path to monetization is murky. Merchants pay for traffic; they pay less for rendering quality without provable lift. You'd need an agent-surface partnership (Anthropic? Perplexity? someone smaller?) to A/B test renderings. Risk: Stripe extends Checkout Studio to cover agent rendering and you're competing with a free feature.

04
Defensibility play

The Japan agent-commerce gateway

Least sexy. Most realistic. Highest defensibility.

Wedge: high TAM: small–medium Defensibility: high

Japan is a massive e-commerce market that is structurally hostile to the current protocol stack. Most merchants are on BASE, Stores.jp, MakeShop, EC-Cube, Rakuten — not Shopify. Payment is fragmented: konbini cash, mobile carrier billing (docomo/au/SoftBank), Pay-easy, JCB, PayPay, LINE Pay. Japanese listings have linguistic quirks (kanji + hiragana + katakana mixing, transliterated brand names, cm sizing, category-specific notation) that English-trained LLMs handle poorly without targeted work.

The American protocol crowd (Stripe, Google, Coinbase) has shallow Japan coverage. Stripe added JP support; Japanese merchants overwhelmingly don't use it.

The gateway, five components

  1. Connectors to BASE / Stores.jp / Rakuten APIs.
  2. Bilingual catalog normalization — JP listing → ACP/UCP schema in EN + JP.
  3. Payment bridging — accept ACP/MPP from foreign agents, settle JPY to JP merchants via Stripe Connect or domestic rails.
  4. Tax / customs / shipping bundling — EMS, Japan Post, Yamato APIs.
  5. Spoken-Japanese agent surface for inbound tourism commerce. "I'm in Shibuya, find the closest store that carries this watch in 38mm and reserve it."

Why you specifically: this is the market Mercari operated in. You know the platforms, the customer behavior, which features actually matter. No American startup will out-execute you here.

Monetization
Gateway SaaS + take-rate on cross-border. Plausibly $1k–$10k MRR per merchant given high cross-border AOV.
The moat
Cultural + linguistic + regulatory. Coinbase isn't building this. Stripe isn't, on your timeframe.

If I had to pick one idea with the highest probability of becoming a sustainable, profitable business for two people with your background — this. Ceiling is lower than Idea 2; floor is much higher.

05
Right idea, wrong year

Returns & disputes for agent commerce

The problem is inevitable. Timing is the question.

Problem will exist: very high Timing: 2027–2028

Stripe Radar covers payment fraud. Nobody yet has an answer for the post-purchase mess agent commerce creates:

The four hairy problems

  1. Agents that bulk-purchase across merchants for comparison, then return 80%.
  2. Agents that hallucinate attributes ("the agent told me it was waterproof but it's water-resistant") leading to disputes about liability — merchant, agent, or agent's operator.
  3. Cross-protocol disputes: bought via ACP, drop-shipped via UCP partner. Who refunds?
  4. Multi-agent transactions (A2A) where the chain of agency makes liability unclear.

A SaaS that integrates with ACP/MPP/UCP transactions, applies an ML returns-risk score at order time, auto-handles agent-triggered returns via standardized workflows, and mediates cross-protocol disputes will be valuable.

Why fifth: timing. The pain is acute in 2027–2028 when agent volumes are real. Today, merchants don't feel it badly enough to pay you. Building product with a few design partners now positions you — just don't expect rapid SaaS revenue for 18–24 months.

§6 What I'd actually do

Combine Idea 1 and Idea 4. Launch bilingual.

If the goal is probability of building a profitable, defensible company in 18–24 months: take the agent-readiness SaaS, bias the GTM hard toward Japan as your launch wedge.

Pitch Japanese merchants on "we make you discoverable to ChatGPT, Claude, and Gemini for international customers." Pitch Western WooCommerce/BigCommerce merchants on "we make you ACP/UCP-ready without rebuilding on Shopify." Use the Agent View Preview (Idea 3's gen-UI feature) as the demo magic that closes sales calls.

It uses every piece of your Mercari DNA — listings, JP commerce, cross-border. It generates SaaS recurring revenue from month two without fundraising. It sets up an option on Idea 2 (your customers are the suppliers, your data is the listing corpus, you can launch a marketplace as Phase 2). Two people ship MVP in 8–12 weeks.

If you'd rather swing

Pick Idea 2 — the agent-native C2C marketplace. Accept higher mortality risk for asymmetric upside. Use your Mercari supply-side relationships if you have them.

What to avoid.

§7 The trap categories
×

x402 facilitator / MCP marketplace

xpay.sh, Coinbase, and a dozen others are already racing here. You have no specific advantage. Confidence: high.

×

Agent wallet plays

Privy, Stripe Link agent wallet, Coinbase Smart Wallet, Crossmint — all well-funded incumbents. Don't play.

×

Generic "agent analytics"

Will be eaten by Stripe Dashboard's agentic surfaces. Console + Dashboard assistant launched at Sessions 2026.

×

Single-protocol commitments

Anything hard-coded to ACP-only or x402-only. The "winner" will be a fragmented mess for years. Build cross-protocol day one.

§8 Tactical

The first 90 days.

Weeks 1–2

Customer development

Five JP merchants on BASE/Stores.jp exporting internationally. Five US WooCommerce merchants doing $1M–$10M/yr. Cold outreach. Interview. Ask what they see and what scares them.

Weeks 3–6

MVP

WooCommerce connector + manual CSV import + LLM normalizer + UCP feed + ACP catalog export + Agent View Preview. Ship behind a free trial for ten design-partner merchants.

Weeks 7–10

Japan unlock

BASE + Stores.jp connectors. "What queries are agents using to find products like yours" report. Start charging design partners. Onboard first paying customers.

Weeks 11–13

Distribution

Three Substack posts: "We A/B tested 50 product titles for agent visibility — here's what worked." Extreme SEO bait. Product Hunt. Target $5–10k MRR by day 90.

Kill criterion: if you don't hit $5k MRR by day 90, you're probably wrong about willingness to pay. Pivot to Japan-only (Idea 4) or move up-stack into merchandising.