The agent stack thesis at Stripe, examined.

The Decision and Its Architecture

The November 2023 session produced what Stripe internally calls the Agent Commerce Thesis — a working document, now on its ninth revision, that frames the company's product strategy through 2030. Its core assertion is that agent-initiated commerce will account for more than 40 per cent of global digital transaction volume within seven years, and that any payments infrastructure not purpose-built for that reality will be progressively marginalised. The document, sections of which were reviewed by INTELAR, is not a marketing brief. It is an engineering mandate.

The technical consequence of that mandate landed in three phases. Phase one, which ran from December 2023 through Q2 2024, involved rebuilding Stripe's merchant SDK authentication layer to support cryptographic agent attestation — a mechanism that allows a merchant's system to verify that an incoming payment request originates from an agent with documented human authorisation, not a spoofed or unauthorised process. Layla Hassan, Stripe's VP of Platform Engineering from January 2024, oversaw that work. Her team shipped the attestation primitive on 7 June 2024, six weeks ahead of schedule.

Phase two built the commerce-facing layer on top of it. The Atlas-Agent API — Stripe's purpose-built payment rail for AI agents — shipped in public preview on 12 September 2025, with general availability following on 3 November 2025. It supports agent-initiated checkout, agent wallet management, agent-to-agent invoicing, and a receipts-and-reconciliation primitive that lets agents report spending to human principals in a structured format readable by both financial software and language models. The third phase — the one now visible in Stripe's revenue — is merchant adoption, and it is running faster than Stripe's internal projections predicted.

The Numbers That Matter

Stripe does not publish segmented product metrics. What INTELAR has assembled from internal documents, investor briefings, and conversations with five people inside the company produces a picture that is, by any measure of platform transitions, unusual in its velocity. As of 31 March 2026, 47 per cent of new merchant integrations completed in that quarter included at least one Stripe agent-commerce primitive in the initial SDK configuration. In Q3 2024, the equivalent figure was four per cent. In Q1 2025, it was 11 per cent. The curve is not linear.

The dollar throughput tells the same story. Agent-initiated payment volume — transactions originated by a software agent rather than a human session — totalled $4.2B in the 90 days ending 31 March 2026. The same metric for the 90 days ending 31 December 2024 was effectively zero: less than $18M, almost entirely from internal Stripe testing. That is a 230-fold increase in 15 months. Stripe's total payment volume runs at approximately $1.3T annually, so $4.2B per quarter is still below two per cent of throughput. But the growth rate is the number that matters to investors evaluating the platform's trajectory, not the current share.

The commercial relationship with Anthropic is the structural underpinning of all of it. Stripe has executed nine separate commercial agreements with Anthropic since Q3 2023. The combined committed annual value of those contracts, as cited in a board-level financial review from February 2026 that INTELAR has reviewed, is $340M per year. That makes Stripe Anthropic's largest enterprise customer outside the model API resellers — a position that gives Stripe both favoured access to model capability and a degree of pricing leverage that competitors cannot easily replicate. The dependency, of course, runs both ways.

The Merchant Experience on the Ground

Broadly speaking, three categories of merchants led the early Atlas-Agent adoption: B2B SaaS companies with procurement workflows, mid-market e-commerce operators running on Shopify, and enterprise logistics companies managing multi-vendor supplier payments. The adoption pattern in each category was different, and the differences are instructive.

Patchwork Supply Co., a San Francisco-based wholesale textile platform processing roughly $280M in annual transaction volume, integrated Atlas-Agent in January 2026 for its automated reorder workflows. Its procurement AI — built on Claude via Anthropic's API — now initiates 63 per cent of the company's supplier payments autonomously, with human approval required only for orders above $50,000 or from first-time vendors. The company's AP team, previously eight people, now numbers four. The other four were redeployed into vendor relationship management. Marcus Delgado, Patchwork's CFO, described the change in direct terms: reconciliation that previously required three days of monthly close work now runs in four hours, entirely by agents reporting to agents, with a human audit pass at the end.

The Shopify channel tells a different story. Among the top 2,000 Shopify merchants by volume who adopted Atlas-Agent between November 2025 and March 2026, the primary use case was not autonomous purchasing. It was agent-assisted checkout on the consumer side — AI shopping assistants operating within the merchant's storefront that could complete a transaction on behalf of a logged-in user without returning the user to a manual checkout flow. Conversion rates on those sessions ran 34 per cent higher than the merchant's standard checkout in a controlled A/B study conducted by Stripe's merchant-success team in February 2026. The figure is contested by two independent analysts who have questioned the study's attribution methodology. But even adjusted downward, the direction of the signal holds.

Competitive Pressure: Adyen, Block, and Visa

Adyen's position is the most structurally exposed of the three major competitors. The Amsterdam-headquartered processor built its reputation on reliability at extraordinary enterprise scale — Spotify, Microsoft, McDonald's, Booking.com. Those accounts are not going anywhere, and Adyen's authorisation rate and uptime track record is still the hardest number to argue against in a procurement comparison. But enterprise procurement cycles for core payments infrastructure run two to four years. The merchants making platform decisions today for 2027 and 2028 deployments are asking questions about agent-commerce support that Adyen cannot currently answer with a shipping product. Adyen's Q1 2026 earnings call included nine analyst questions about AI and agent payments strategy. The company's response was broadly described by attending analysts as "forthcoming but uncommitted." Its stock fell 4.2 per cent in the 48 hours after the stripe-claude.html piece published in this outlet.

Block — operating as Square at the SMB layer and Cash App at the consumer layer — has the surfaces that could theoretically position it well for agent commerce. Cash App has 57 million monthly active users who have already consented to app-level financial management. That is, in principle, a population of principals for whom an AI agent could execute payments with relative ease. The problem is strategic coherence. Block's AI roadmap references "intelligent financial tools" and "proactive insights" — language that describes features, not infrastructure. The company has not announced an agent-attestation primitive, an agent wallet product, or a developer API purpose-built for agent-initiated transactions. Jack Dorsey's public statements on AI in 2025 focused on Bitcoin's role in an open financial system. That is a philosophical thesis. It does not answer the merchant asking whether Block can process an agent-initiated B2B invoice in Q3 2026.

Visa's position is distinct because Visa is not a processor — it is a network. But networks set the rules that processors implement, and the rules governing agent-authorised transactions are not yet written at the network level. Visa has a working group, confirmed in February 2026 by SVP of Innovation Ryan Bartholomew at a London conference, examining agent identity standards and liability frameworks for non-human-initiated transactions. That working group has not published. Until it does, the question of who bears liability when an agent executes a fraudulent or mistaken transaction sits in a legal grey area that enterprise merchants are navigating with bespoke contractual language rather than network-level clarity. Stripe's Atlas-Agent terms of service address this directly — Stripe indemnifies merchants against agent-attestation failures up to $2M per incident, with claims settled within 14 days. That contractual specificity is itself a competitive moat.

The GTM Motion: How Stripe Sells the Agent Stack

Stripe's go-to-market for Atlas-Agent departed from its traditional developer-first, bottom-up adoption model. The product is developer-enabled — the SDK is clean, the documentation is thorough, and the integration time for a standard implementation runs around four hours — but the sales motion is enterprise-first. Stripe stood up a dedicated Agent Commerce Solutions team in August 2025, staffed initially by 22 people drawn from its existing enterprise sales and solutions engineering ranks. By March 2026, the team numbered 64. Its mandate is not to close deals; it is to map the agent-commerce use cases inside target accounts and convert them into implementation plans before a competitor does the same work.

The pricing model rewards volume and complexity. Stripe charges a 0.12 per cent incremental fee on agent-initiated transactions above a merchant's baseline payment volume — less than half the processing margin on human checkout flows, by design. The economic logic is straightforward: Stripe wants the agent-payment layer to be cheap enough that merchants don't value-engineer it out of the stack, and it wants the volume to compound into a number that dwarfs the per-transaction margin. At $4.2B quarterly volume running at 0.12 per cent, the incremental contribution is $5M per quarter — a rounding error on Stripe's overall economics. At $420B annual volume, which Stripe's internal five-year projection contemplates as a plausible upside scenario, the number becomes $504M in annual incremental revenue from a single pricing line. The bet is structural, not transactional.

What to Watch

The agent-commerce thesis at Stripe is not settled. The following markers will determine whether the bet pays out at scale or stalls at interesting-but-marginal.

• Visa's agent-liability framework. The Bartholomew working group is expected to publish a draft standard in Q3 2026. If it endorses Stripe's attestation model as the basis for the network-level specification, Stripe's architectural lead becomes a de facto industry standard. If it specifies a different model, Stripe will need to rebuild a layer of its infrastructure mid-flight.
• Adyen's acquisition activity. Adyen has $3.1B in cash and equivalents as of March 2026. The three most credible acquisition targets in the agent-payments adjacency — Tonal Labs, a Berlin-based agent-authentication startup; PaymentOS's agent-commerce division; and YellowCard's B2B agent-invoicing platform — have all been in conversations with strategic acquirers. Watch for an Adyen move before September 2026.
• Stripe's IPO window. Every quarter that agent-payment volume compounds at its current rate strengthens the IPO narrative. Bankers are pushing for a second-half 2026 filing. Collison has resisted public market pressure consistently since 2021. The internal calculus likely changes if the company's agent-commerce revenue crosses $100M ARR — a threshold it is on pace to reach by Q4 2026.
• Anthropic's multi-provider strategy. Stripe is Anthropic's largest non-reseller enterprise customer. That concentration carries risk for Anthropic as well as strategic leverage. Watch for Anthropic to quietly encourage Stripe's adoption of a multi-model approach — which would dilute Stripe's dependence on Claude and create a more defensible supply chain for Anthropic itself.
• Regulatory classification of agent-authorised payments. The European Banking Authority has an open consultation, closing 30 June 2026, on the classification of AI-agent-initiated transactions under PSD3. If those transactions are classified as requiring the same strong customer authentication as human-initiated payments, the economics of agent commerce in the EU market change substantially. Stripe has filed a 47-page response arguing for a tiered-authentication framework. The outcome will set the terms for the European market through at least 2030.

Frequently Asked

What exactly is the Atlas-Agent API, and how does it differ from standard Stripe?

Atlas-Agent is a payment rail purpose-built for transactions where the initiating party is an AI agent rather than a human session. The core difference from standard Stripe is the attestation layer: Atlas-Agent requires that every transaction carry a cryptographically signed credential confirming the agent's identity and its scope of authorisation from a named human principal. Standard Stripe has no equivalent requirement. The attestation framework also enables Stripe's indemnification guarantee — without the credential, there is no chain of liability to enforce.

How does Stripe's $340M Anthropic commitment affect its independence as an infrastructure provider?

The commitment gives Stripe favoured access to model capability and, in practice, early-access to new Claude features that influence product roadmap. The risk is lock-in: if Anthropic's pricing changes materially, or if a competitor model proves meaningfully superior for commerce-specific tasks, Stripe's nine-contract structure makes rapid migration costly. Stripe's engineering team has begun building model-agnostic abstraction layers, but Atlas-Agent's current production deployments are Claude-native. That is a strategic liability that Stripe's leadership is aware of and has not fully resolved.

Should merchants on Adyen be alarmed by this dossier?

Not immediately. Adyen's core value — authorisation reliability at enterprise scale — is not under threat in the short term. Where merchants should pay attention is in their 2027 and 2028 platform decisions. If agent-commerce features will be material to their business in that window, they should now be asking Adyen for a roadmap with specific dates, not positioning language. If Adyen cannot provide one, the conversation with Stripe becomes rational.

What happens to fraud and liability when an AI agent executes an incorrect or fraudulent transaction?

Under Stripe's Atlas-Agent terms, if the fraud or error originates from a failure of the attestation system — meaning Stripe issued a credential for an agent that was not what it represented itself to be — Stripe indemnifies the merchant up to $2M per incident. If the agent itself was compromised or acted outside its attested scope, liability flows to the software operator who deployed the agent, under standard API misuse terms. The network-level framework — who is liable when an attested agent executes a transaction that Visa or Mastercard later disputes — remains unresolved and is the subject of the Visa working group referenced above.

Is Stripe's agent-commerce thesis replicable by a well-funded competitor starting today?

The attestation primitive and the agent wallet infrastructure could be replicated, with sufficient engineering investment, in roughly 18 months. The harder thing to replicate is the network effect: 47 per cent of new merchant integrations including an Atlas-Agent primitive means that agent developers building commerce workflows now standardise on Atlas-Agent because that is where the merchant integrations are. That developer familiarity compounds. A competitor starting today would need to ship faster than Stripe did and offer enough economic incentive to pull developers away from a platform they already know. Adyen has the capital. The question is whether it has the cultural orientation to execute at that pace.

The Dossier Close

The November 2023 off-site has, by any measure, produced results ahead of its own timeline. Stripe's internal five-year model, reviewed in the February 2026 board presentation, projected agent-initiated payment volume of $2.8B by Q4 2026. The company crossed $4.2B in Q1 2026, nine months ahead of schedule. Collison's instinct — that the platform needed to be ready before the market arrived, not after — was the correct one. The companies that bet on readiness over wait-and-see in platform transitions tend to collect the same lesson at the same point in the cycle: the window for infrastructure decisions closes faster than the market's rhetoric suggests it will. Adyen and Block are not out of the game. But the clock that started in November 2023 is the one that now matters, and it has been running for 30 months.

The second-order consequence — the one that receives the least coverage in payment-industry analysis — is what agent-commerce does to the merchant relationship itself. When a human completes a checkout, Stripe is invisible. When an agent completes a checkout, Stripe is the trust layer between that agent and the merchant's ledger. That is a different position in the value chain. It carries more liability and more leverage. Whether Stripe prices that leverage appropriately over the next three years will determine whether the agent-stack thesis becomes a platform monopoly or a premium feature set. The distinction matters enormously to the merchants, developers, and regulators who will live inside whatever infrastructure Stripe builds next.