Stripe's 5 Levels of Agentic Commerce
Stripe’s 2025 letter takes “agentic commerce” out of sci-fi land and turns it into a staircase. Five distinct steps, each one closer to the thing everyone is vaguely pointing at.
From ATXP’s perspective, that framing is mostly right — while still undershooting the weirdness of what’s coming. So this piece is effectively “yes, and…” the most important shift isn’t agents buying things. It’s the internet learning how to accept intent as a first-class object, with payments as the enforcement mechanism.
To recap, Stripe’s five levels of agentic commerce are:
- eliminate web forms
- descriptive search
- persistence
- delegation
- anticipation
We agree with the “small chunks” approach Stripe explicitly prefers. The hype version skips all the hard parts. Where we diverge is what we think each step actually requires underneath — and what the “missing steps” are between them.
*Disclosure: we worked at Stripe, they invested in us, and they didn’t know we were writing this piece. A fun recipe!
Level 1: Eliminating web forms (it’s a cul-de-sac, not an exit ramp)
Stripe describes Level 1 as “typing and clicking ‘buy’ on your behalf,” with the human still doing all the deciding.
Yes. And this level is mostly an interface hack. It’s valuable (forms are terrible), but it doesn’t change the substrate. It’s screen-scraping with better manners.
The risk is that teams mistake “we automated checkout” for “we built agentic commerce.” Level 1 is a bridge — not a destination — because it scales poorly the moment you want reliability, auditability, or anything that looks like a business process rather than a one-off purchase.
ATXP view: Level 1 is a migration path for incumbents, not the end state for new systems.
Level 2: Descriptive search (first real unlock)
Stripe’s Level 2 is where commerce stops being keyword search and becomes “describe your situation,” with the system reasoning across constraints like price, durability, delivery timelines, taste, etc.
This is the first step that actually changes the unit of value. A product catalog isn’t the thing anymore; the thing is an optimizer over constraints.
Yes. And descriptive search is not primarily a discovery upgrade — it’s a specification upgrade. It teaches human users to express intent in a machine-usable way. That matters because it becomes the input to delegation later.
If Level 2 wins, commerce starts to look less like “shopping” and more like “curating.”
Level 3: Persistence (memory isn’t the hard part — permission is)
Stripe frames Level 3 as: you stop reintroducing yourself; the agent remembers preferences inferred from prior conversations and purchases; you still decide what to buy.
Yes. And the bottleneck isn’t “can the model remember.” It’s “what is the scope of that memory, who owns it, and how does it travel?”
Persistence without portability becomes a walled garden story by default. Most companies want to own a walled garden. We side with Stripe, wanting the opposite — universal interoperability, and they’re pushing open-ish plumbing (ACP) to get there.
ATXP view: persistence will be less about one agent having a perfect profile of you, and more about selectively provable traits you can lend to agents for a job:
- “this purchase is within my budget policy”
- “this merchant is in my allowlist”
- “this item matches my sizing constraints”
- “this vendor meets my return/shipping requirements”
In other words: memory is useful, but permissioning is destiny.
Level 4: Delegation(?)
Stripe’s Level 4 is “you stop choosing altogether,” you provide a budget, and the system searches, evaluates, and purchases on your behalf. This is where most people tend to land when they envision “agentic commerce.” It’s well positioned at Level 4 instead of Level 1.
Yes. And delegation only works at scale if it’s reversible and inspectable.
Humans don’t trust autonomy; they trust:
- clear budgets
- constrained authority
- receipts
- and the ability to unwind mistakes
So between Level 3 and Level 4, there’s an implied “Level 3.5” Stripe doesn’t name: policy + audit. Ramp has built a massive business by helping businesses look beyond 2% cash back, when they could save the other 98% by not spending in the first place.
ATXP view: the winning primitives here look like:
- spend limits as programmable policies
- “up to $400, but also: no third-party marketplaces, must arrive by Friday, prefer merchants with easy returns”
- “If a $150k employee saves two hours late on a Wednesday by having dinner delivered, approve the $13 burrito”
- cryptographic or at least verifiable “who authorized what” or “what was the exception”
- machine-readable invoices/receipts and reason traces (“I chose these items because…”)
Delegation is not one big “yes.” It’s lots of small “yeses” that are easy to refine. And revoke!
Level 5: Anticipation (agree it’s futuristic, but “no prompt” is misleading)
Stripe’s Level 5 is anticipation: “there is no prompt,” the system already knows the calendar, preferences, budget, and purchases things before you ask. This is directionally right, but I’d rephrase it:
The prompt is always there; but you might not be best positioned to write it. Sometimes your friends can hold up a mirror and reveal things you can’t see about yourself. And – while often painful! – that can be hugely impactful.
Calendar events, inventory levels, household routines, business workflows, implicit SLAs. Anticipation is what happens when agents can read those signals and have permission to act. Think about how you manage your calendar, how many new years resolutions you stick to, where you waste time vs where you lose track of it. Hopefully level 5 unlocks more of the latter.
Dialing back the philosophizing, Stripe is really saying: the latter levels won’t happen without shared rails.
The real thesis: Interoperability beats magic
Stripe explicitly compares the current moment to the mid-90s protocol era and argues the future depends on universal interoperability. And they’re building toward it with:
- Agentic Commerce Protocol (ACP), with OpenAI, as shared language
- Shared Payment Tokens to let agents initiate payments without exposing credentials
- Agentic Commerce Suite to sell across multiple AI interfaces/protocols via one integration
- “machine payments” for charging agents for API calls/MCP/HTTP via stablecoin micropayments
We like this a lot. It’s the right instinct: don’t bet on a single agent UI; bet on the underlying payment + identity + permission primitives that let many interfaces coexist.
ATXP’s “yes, and” is that machine payments isn’t a bullet point — it’s the seed of an entirely different market structure.
Once agents are customers, pricing stops being “$20/mo per seat” and starts being:
- per call
- per outcome
- per delegated action
- per verified task completion
You don’t just monetize humans more efficiently. You get brand new demand from software that can pay.
Where we disagree (gently): the staircase is missing a basement
The five levels describe what users experience. But there’s a lower layer (Level 0? Level -1?) that determines whether the whole thing becomes:
- a set of walled gardens with good UX, or
- an open economy where agents can roam and transact safely.
That basement layer is:
Identity, permission, and payments as composable infrastructure for non-human actors.
Stripe is clearly building pieces of it (tokens, ACP, machine payments). We think the opportunity is bigger: a world where “who can do what, with which funds, under what constraints” is programmable and portable across agents, merchants, and tools.
If the early internet was “information wants to be linkable,” the agentic internet is “intent wants to be routable.”
(If that sounds like a small semantic shift: it isn’t.)
Agentic commerce is commerce where an AI agent — acting with delegated authority — initiates, evaluates, and completes transactions on a user's behalf. It spans a spectrum from simple form automation to fully anticipatory purchasing, with each level requiring progressively more trust infrastructure: identity, permissioning, and payment primitives that work for non-human actors.
npx atxpBuild agents that can participate in agentic commerce — with identity, payments, and policy controls built in. The staircase is missing a basement → · Every agent payment protocol compared →
Frequently asked questions
What are Stripe’s 5 levels of agentic commerce?
Stripe’s framework describes five steps: (1) eliminating web forms, (2) descriptive/natural language search, (3) persistence and preference memory, (4) delegation where agents act autonomously within a budget, and (5) anticipation where commerce happens before the user prompts it.
What is the “missing basement” ATXP identifies below Stripe’s framework?
The missing basement is identity, permission, and payments as composable infrastructure for non-human actors. Stripe describes the user experience at each level but doesn’t fully address the underlying layer that determines whether the result is open interoperability or a set of walled gardens.
Why does ATXP argue that “permissioning is destiny” at Level 3?
Persistence without portability defaults to walled gardens. The critical question isn’t whether agents can remember preferences — it’s who owns that memory, what scope it covers, and whether it can travel across services. Selectively provable traits (budget policy, merchant allowlist, sizing constraints) are more powerful than a monolithic profile.
What does “intent wants to be routable” mean?
It’s a reframe of “information wants to be free.” In the agentic internet, the primitive isn’t a URL or a document — it’s a statement of intent with accompanying permissions and payment authority. Routing that intent to the right merchant or service, reliably and verifiably, is the core infrastructure problem.
How does ATXP relate to the protocols Stripe is building (ACP, machine payments)?
ATXP sits at the function-level billing layer above Stripe’s transport and wallet primitives. ACP handles agent-to-agent permissions; machine payments handle HTTP-level settlement; ATXP handles per-tool pricing and usage controls. These layers complement rather than compete.
Is Level 4 delegation possible without Level 3.5 policy and audit?
Not reliably. ATXP identifies an implied Level 3.5 — policy plus audit — that Stripe doesn’t name explicitly. Humans trust delegation through clear budgets, constrained authority, receipts, and reversibility. Skipping that layer means delegation without accountability.