When people imagine agentic commerce, they picture the climax: an agent, acting for you, completing a purchase. So the obvious gold rush is the checkout — protocols, buttons, the moment of transaction. I think that is the least interesting part of the whole system, and probably the least defensible.
The transaction is the easy mile
Pressing “buy” is a solved problem dressed up as a hard one. Payment rails exist. Tokenization exists. The hard part was never the final tap; it was everything that has to be true before an agent can responsibly take that action on someone’s behalf.
Agentic checkout is probably not the opportunity. The opportunity is everything that has to be true before checkout is safe.
Where the real work lives
Walk backward from the transaction and the genuinely hard problems appear:
- Catalog truth. An agent acting on bad product data makes confident, wrong purchases. Whoever maintains a clean, machine-legible, continuously-correct view of what is actually for sale holds something durable.
- Trust and constraints. “Buy the thing I’d buy” encodes a person’s taste, budget, and red lines. Capturing and honoring those constraints is a product, not a setting.
- The merchant’s side. Every buyer agent implies a seller surface. Merchants need to expose inventory, policy, and intent in a form agents can reason about — and they need to trust the agents they’re transacting with.
- Reconciliation and recourse. When an autonomous purchase goes wrong, the system that handles the return, the dispute, the apology is where loyalty is won or lost.
Sell shovels to the part that’s hard
The pattern repeats across every automation wave: the visible, demo-friendly step gets commoditized fastest because it is the easiest to copy and the easiest to standardize. Value accrues to the unglamorous infrastructure that makes the visible step trustworthy.
For commerce specifically, that means the opportunity is less “agent that shops for you” and more the layer that lets any agent shop safely: verified catalogs, portable preference, merchant-side intent, and the boring machinery of trust. It is harder to build and harder to demo. That is the point. The thing that is hard to demo is the thing that is hard to copy.
If you are building here, resist the pull toward the checkout. The checkout will be a feature in everyone’s stack within a year. Go find the part beneath the surface that has to be true for the checkout to mean anything.