Stripe’s 288-Feature Launch: Agent Commerce Is a Plumbing Problem, Not a Model Problem

Agent commerce is a payment problem now

At Stripe Sessions, the company released 288 new features and introduced an agent wallet built on Link. The headline is not the count; it is the framing. Stripe is no longer pitching AI as a feature for checkout buttons. It is positioning itself as the transaction layer for a machine-run economy.

The wallet is not merely a new payment method. It is an attempt to give agents the same kind of credential that human cardholders already have: a stored identity, trusted by merchants and issuers, capable of initiating a transaction without asking for permission every time.

Agents are the next counterparty, not the next interface

Patrick Collison argued that agents may soon initiate most transactions. That changes the job of a payment platform. Today, the buyer is a human with a device, an identity, a credit limit, and a dispute history. Tomorrow, the buyer may be an autonomous process acting on a budget, a scoped mandate, and an audit trail.

This is different from conversational commerce. A voice assistant can recommend a product; an agent can book it, cancel it, and reorder it. The distinction is not chat; it is state change. The payment challenge is not whether the agent can speak, but whether the system can let it spend without letting it run wild.

288 features, one plumbing agenda

Stripe’s 288-feature release is best read as a long list of prerequisites for agentic commerce. The agent wallet gives a machine actor a stored payment identity, but identity is only the beginning. The real work is authorization: per-transaction limits, merchant allowlists, reversible permissions, and logs that can be inspected after something goes wrong.

Link is a useful starting point. Stripe already operates a network of stored consumer credentials. Repackaging that infrastructure for agents means the network effect can be reused, but it also means the trust assumptions must be rebuilt around machine principals rather than human cardholders.

When the agent buys the wrong flight or a merchant disappears, the dispute path must be as automated as the purchase path. That means turning the human-read, human-reviewed parts of the payment back end into machine-readable policies and rules. Otherwise, every agent purchase becomes an opportunity for manual firefighting.

• Identity: an agent must prove which principal it represents and which wallet it may draw from.
• Authorization: the principal must be able to set granular constraints on what the agent can buy, from whom, and up to what amount.
• Audit: every action must leave a trail that maps back to a human or organizational owner.

When the risk surface is not eloquence but context

The adjacent news in early May makes this risk more concrete. The UK AI Security Institute showed GPT-5.5 completing multi-step cyber attack simulations, proving that frontier models can execute chains of actions, not just answer questions. Anthropic’s Claude Security now scans code and proposes fixes inside enterprise workflows, raising questions about who approves the patch and who owns the error.

Meanwhile, Claude Code was found to be influenced by stray JSON in commit history, a reminder that an agent does not read only the prompt; it reads the whole messy context. These are not isolated security bugs. They are previews of the authorization problem Stripe is trying to solve in the narrower domain of money.

Developer experience is now agent experience

Collison also noted that AI makes developer experience more strategic. The reason is simple: agents consume APIs at scale. A confusing endpoint, an ambiguous error code, or a permission model that requires human consent at every step will break an agent workflow.

The platforms that win will be those that expose clean, deterministic, reversible operations. In an agent context, an ambiguous API response is not just a UX issue; it is an audit failure waiting to happen. Eloquence may be the product demos, but determinism is the infrastructure.

The moat is trust, not token count

AI companies are being priced like infrastructure assets, yet the infrastructure they need is not all theirs to build. Models can be swapped; a relationship with merchants, banks, card networks, and regulators cannot. If agents become the default initiators of transactions, the durable advantage will belong to whoever can answer the boring questions: who authorized this, where is the money, how do we reverse it, and what is the record?

Stripe’s 288-feature launch is a bid to own that layer. The bet is that agentic commerce’s next bottleneck is not the model’s vocabulary, but the plumbing of payment, identity, and authorization.