Why we run one agent, not a marketplace of agents

The obvious architecture for an AI-commerce platform is a marketplace: let creators pick from competing agents, may the best one win. We considered it seriously. We built the other thing instead — one engine, gmvagent, running every store — and this essay is the reasoning.

Improvements compound. When every store runs on the same engine, a fix ships to everyone at once. Better mockup rendering, a smarter escalation rule, a fulfilment edge case handled — thirty-plus stores get it the same day. In a marketplace, every improvement is trapped inside one vendor's product.

Trust becomes auditable. A buyer talking to any Pinpinmo storefront is talking to the same engine under the same rules: no invented discounts, no over-promised delivery dates, money decisions escalated to humans. We can state that as a platform guarantee precisely because there's one system to hold to it. With N vendors, the guarantee dissolves into N privacy policies.

Creators keep the control that matters. Uniform engine does not mean uniform stores. Every store has its own brand voice, its own catalog, its own pricing, and its own guardrail settings deciding what the agent may do alone. The dial creators actually want — how much autonomy, where — survives. The dial they don't want — which vendor's black box to bet the store on — is gone.

The escape hatch is real. One engine doesn't mean closed. The Agent API lets any creator wire in their own automation with scoped keys, the same permission checks, and the same audit log. You can bring your own brain; it just has to play by the store's rules.

A marketplace optimizes for choice. A platform optimizes for the floor never dropping. For commerce — where someone's money is always on the table — we'll take the floor.