The AI Agent Payments Landscape in 2026 - Protocols, Players, and Primitives

The shift from Card-on-File to Agent-on-File is happening faster than anyone predicted.

In December 2025, Visa announced that hundreds of secure, agent-initiated transactions had been completed across its partner ecosystem. The company predicted millions of consumers would use AI agents to complete purchases by the 2026 holiday season. Mastercard rolled out Agent Pay to all U.S. cardholders by November 2025. Stripe and OpenAI published the Agentic Commerce Protocol as an open standard.

AI agents are no longer just chatbots answering questions. They're becoming economic actors with wallets, spending limits, and transaction histories. The infrastructure supporting this transition has exploded into a complex ecosystem of wallet providers, identity layers, card issuers, execution platforms, and network standards.

This post maps the landscape. We'll cover who's building what, how the pieces fit together, and where the industry is headed. The goal isn't to declare winners but to help anyone researching options understand the current state of play.

The Paradigm Shift: Agents as Economic Actors

Traditional ecommerce assumes a human behind every transaction. You browse a website, add items to a cart, enter your card details, and click purchase. Fraud detection systems look for human behavior patterns. Merchant checkouts are designed for human eyes.

Agentic commerce inverts this model. An AI agent researches products, compares prices, navigates checkout flows, and initiates payment, all without human intervention for each transaction. The human's role shifts from executing purchases to setting policies and approving exceptions.

This shift creates new infrastructure requirements. Agents need identity credentials that merchants can verify. They need payment instruments with programmatic controls. They need execution environments that can handle browser automation. And they need protocols that let disparate systems communicate securely.

The companies building this infrastructure fall into several distinct layers.

Infrastructure and Wallet Providers

The foundation of agentic payments is the ability for an agent to hold value and initiate transactions. Several companies are building this infrastructure layer with different architectural approaches.

PayOS emerged from stealth in April 2025 with partnerships across both Visa and Mastercard. Their approach centers on network-issued Agent Tokens, secure credentials that enable AI-driven transactions at any merchant accepting card payments. PayOS handles the complexity of integrating with payment networks while providing APIs that developers can build against.

The company has positioned itself as processor-agnostic, meaning developers can use PayOS regardless of their existing payment infrastructure. Their business model includes facilitation fees and tools for AI usage billing. In September 2025, PayOS demonstrated the first live transaction using a Mastercard Agentic Token in a controlled proof-of-concept, marking an early milestone for network-rail agentic payments.

PayOS's token-based approach represents one side of a fundamental architectural debate. Token delegation offers broad merchant acceptance and network-level integration. The tradeoff is that shared credentials create blast radius risk: if a token is compromised or an agent misbehaves, the exposure extends to any merchant where that token is valid. For enterprises prioritizing isolation over flexibility, dedicated card approaches may be preferable.

Nekuda takes a different approach built around what they call the Mandate Model. Where PayOS focuses on tokens, Nekuda focuses on capturing user intent. Their system creates "agentic mandates" that specify what an agent is allowed to buy, under which conditions, with what spending limits, and when approvals are required.

Nekuda raised $5 million in May 2025 from Madrona, Amex Ventures, and Visa Ventures. They're a launch partner for Visa Intelligent Commerce and have completed pilots enabling AI agents to move from product discovery to purchase. Their promise of "20 lines of code" integration appeals to developers who want fast time-to-market.

Prava focuses on multi-agent deployments with a single wallet infrastructure. Their API lets developers give agents prepaid wallets separate from bank accounts or credit cards. One developer wallet can fund multiple agents, each with its own spending rules. This architecture suits platforms running many agents simultaneously, like customer service deployments or automated procurement systems.

Proxy takes yet another approach: dedicated virtual cards for each agent or workflow. Rather than shared tokens or mandates, each agent gets its own card with hard spending limits enforced at the network level. The philosophy is attestation-before-access, meaning agents must prove their authorization before any transaction attempt.

The tradeoff is operational overhead. Managing individual cards is more complex than shared credentials. But the isolation provides stronger blast radius control and simpler audit trails when something goes wrong.

Each approach reflects different assumptions about how agentic commerce will evolve. Token-based systems optimize for flexibility and network integration. Mandate systems optimize for capturing user intent. Dedicated cards optimize for isolation and control.

Identity and Trust Layer

Before an agent can transact, merchants need to know they're dealing with a legitimate actor, not a bot, fraud attempt, or competitor scraping prices. The identity and trust layer addresses this challenge.

Skyfire has built the most comprehensive identity system for AI agents through their KYAPay protocol. KYA stands for "Know Your Agent," and the protocol lets businesses identify and verify agents attempting to access their services.

The technical implementation uses signed JSON Web Tokens that carry verified information about the agent owner, authorized spend amounts, and payment credentials. Merchants can verify agent identity without exposing underlying credentials. The system supports building a verifiable track record of agent activity over time, creating something like a credit history for AI agents.

Skyfire also provides enterprise controls including just-in-time decisioning, where humans can approve high-value transactions via SMS before they execute. The company is backed by a16z CSX, Coinbase Ventures, and others, with a focus on making agents "full participants in the digital economy."

Visa's Trusted Agent Protocol (TAP) brings identity to the network level. Launched in October 2025, TAP adds digital proof-of-identity to agent-initiated transactions using the Web Bot Auth standard for HTTP message signatures.

The technical mechanics work as follows: agents must register their public keys in a Visa-managed directory before initiating transactions. When an agent makes a request, it cryptographically signs the HTTP message using its private key. Merchants and Visa can then verify the signature against the registered public key, confirming that "this agent is authorized to act on behalf of cardholder X." This verification layer works alongside Visa's network tokenization, adding agent identity on top of existing card security.

For an agent to use TAP, it must first be approved through Visa's Intelligent Commerce vetting program. Approved agents receive unique cryptographic keys that serve as their identity credentials. The protocol includes three components: Agent Intent (proving the agent is trusted and specifying its purpose), Consumer Recognition (indicating whether the consumer has an existing relationship with the merchant), and Payment Information.

TAP was developed in collaboration with Cloudflare and is available on GitHub. Akamai announced integration with TAP, combining Visa's identity layer with edge-based behavioral intelligence and bot protection. This lets merchants distinguish trusted AI agents from malicious bots.

The Agentic Commerce Consortium represents an industry-wide effort at standardization. Formed in September 2025 by Basis Theory, the consortium includes Lithic, Skyfire, Rye, Crossmint, and others. They published a white paper titled "Empowering Merchant-Controlled AI Commerce" that lays groundwork for how merchants can opt into agentic commerce on their own terms.

The consortium is solving problems across the stack: authorization and verification, product catalog discovery, website agentification, unified shopping cart APIs, and payments. Basis Theory raised $33 million in October 2025 to continue this work, with their programmable vault serving as foundation infrastructure for the consortium's standards.

Card Issuance

The issuance layer provides the actual payment instruments that agents use. These companies offer APIs for creating virtual cards programmatically with custom authorization rules.

Lithic has positioned itself as purpose-built for agentic commerce. Their Authorization Stream Access (ASA) feature lets developers build unique rules for each card, evaluated in real-time as transactions arrive. This enables what they call "just-in-time auth," where agents receive payment permissions precisely when needed and scoped to exact merchant and amount.

Lithic partnered with Arcade.dev in August 2025 to build end-to-end agentic commerce workflows. The combination of Lithic's programmable cards with Arcade's tool-calling framework and Rye's checkout API demonstrated that the pieces could work together in production. The company is also a founding member of the Agentic Commerce Consortium.

Marqeta brings enterprise scale to agentic payments. Their Model Context Protocol (MCP) Server, built on Anthropic's open standard, lets AI agents directly access Marqeta's payment APIs. With over $200 billion in annual Total Payment Volume and more than a billion dollars processed in a single day at peak, Marqeta handles transaction volumes that few others can match.

Their MCP Server enables agents to provision virtual cards instantly, operate within defined policies, create and submit disputes, and retrieve transaction details with anomaly flagging. The company appointed a Chief AI Officer in 2025, signaling commitment to AI-driven innovation.

Highnote differentiates through stablecoin settlement capabilities. In December 2025, they launched a stablecoin settlement pilot with Cross River Bank and Visa, using USDC over Solana for production settlement. This reduces settlement from 2-3 business days to minutes and enables 24/7 availability.

Earlier in 2025, Highnote partnered with BVNK for real-time stablecoin-based funding, letting global customers fund U.S. program accounts instantly. For high-velocity use cases like supplier payouts or expense reimbursements, the speed advantage compounds.

Rain operates at the intersection of stablecoins and card rails. As a Visa principal member, Rain issues cards backed by on-chain stablecoin wallets. This lets businesses hold self-custodied stablecoins while transacting at any Visa merchant.

Rain joined Visa's stablecoin settlement pilot, fully tokenizing credit card receivables and settling all Visa transactions in USDC. They've also launched what they call "closed loop credit card receivable financing" using stablecoins, reducing total cost of capital for credit programs. The company has processed payments in over 150 countries.

Execution Layer

Having a wallet and card isn't enough. Agents need to actually navigate websites, fill forms, and complete checkouts. The execution layer handles this browser automation challenge.

Rye provides the Universal Checkout API that lets AI shopping applications complete purchases from any online store using just a product URL. The technical challenge they solve is significant: merchant fraud detection systems have blocked automated transactions for decades.

Rye's system combines browser automation with fraud mitigation techniques. Their Merchant Risk Adapter routes interactions through residential, geo-proximal IPs matching the buyer's shipping region. They tune browser parameters and interaction profiles to resemble human behavior. Page timings, scroll patterns, mouse paths, and form-filling cadences all mimic natural usage.

When an automated flow works, Rye caches it as a deterministic workflow for speed. When layouts change, self-healing fallbacks kick in. The company claims offer resolution in under 35 seconds and checkout completion in under 10, with reliability above 90% across major platforms.

Rye originally raised $14 million from a16z crypto in 2022 with a vision of making ecommerce programmable. By 2024, they pivoted to become "the checkout engine every AI agent would need."

Induced AI provides cloud browser environments specifically designed for AI agents. Their platform runs browser-based tasks 24/7 with built-in anti-bot detection avoidance, secure authentication, and real-time reasoning capabilities.

The company targets enterprises looking to automate repetitive browser tasks like customer service, data entry, and online research. Multiple parallel workflows can run without getting blocked, and disposable accounts provide secure access to various platforms.

By mid-2025, "agentic browsers" became a category unto themselves. Tools from Perplexity (Comet), Browser Company (Dia), Microsoft (Copilot in Edge), and others reframed browsers as active participants rather than passive interfaces. Anthropic's Model Context Protocol, released in late 2024, provided the standard for connecting language models to external tools, effectively enabling browsers to act on model instructions.

B2B and Corporate

Enterprise use cases have their own requirements around compliance, approvals, and integration with existing financial systems. Several companies target this segment.

Natural.co launched in October 2025 with $9.8 million to build payments infrastructure for AI agents in B2B contexts. Their focus is on industries where agents can source, negotiate, and pay vendors in real time: logistics, property management, procurement, healthcare, and construction.

The vision includes agents managing procurement and inventory, automating vendor payments in healthcare, hiring and paying contractors in construction, and negotiating freight settlements in logistics. Natural is keeping the team lean, expecting no more than 10 employees by end of year, and plans general availability in 2026.

Payman AI addresses a different problem: what happens when AI agents need human help? Their platform creates a marketplace where agents can pay humans for tasks they cannot complete themselves.

When an AI encounters complex work requiring judgment, like code reviews, design input, or legal analysis, it can post the task to Payman and pay a human expert to complete it. The platform enforces spending limits and human approvals, with policies that trigger manual sign-off when transaction sizes exceed thresholds.

Payman is built on SOC-2 and PCI controls with Fifth Third Bank for USD custody. Use cases span design, code review, gaming, medical diagnosis support, sales execution, and marketing.

Brex announced a partnership with Fifth Third Bank in December 2025 that unlocks $5.6 billion in annual commercial card volume. The Fifth Third Commercial Card powered by Brex will become the default solution for the bank's commercial clients.

The AI agents built into Brex handle expense management end-to-end: an expense agent grabs receipts, a review agent checks policy compliance, an auditing agent eliminates manual review, and an accounting agent syncs charges to ERP systems. Brex claims these agents automate 85% of expense reviews with 99% accuracy and catch 15x more out-of-policy spend than traditional rules-based flags.

Ramp has deployed specialized AI agents for controllers that automatically enforce expense policies, eliminate unauthorized spending, and prevent fraud. Built on OpenAI's reasoning models, the agents apply context-aware judgment to manage workflows independently.

Ramp's Agents for AP automate invoice coding, approval, and payment processing. Early access customers reported $1 million in fraudulent invoices flagged in 90 days. The company was used by over 45,000 businesses as of October 2025 and claims to have saved customers $10 billion and 27.5 million hours of labor.

Network Standards and Protocols

The major payment networks and platforms have all published protocols for agentic commerce. These standards will shape how the ecosystem develops.

Visa Intelligent Commerce is the umbrella initiative encompassing TAP and other agentic capabilities. Visa is working with more than 100 partners worldwide, with over 30 actively building in their sandbox and over 20 integrating directly. The company expanded Intelligent Commerce globally in late 2025, with pilots launching in Asia Pacific, Europe, and Latin America in early 2026.

Mastercard Agent Pay uses Agentic Tokens, dynamic digital credentials that bring tokenization technology from mobile payments to agentic commerce. Each agent is uniquely identified and enabled to initiate transactions using cryptographically secure credentials.

Mastercard has partnered with Microsoft on new use cases and Cloudflare for Web Bot Auth support. Their Agent Pay integrates with Google's Universal Commerce Protocol and PayPal's wallet. Citi and US Bank cardholders were enabled by late 2025, with global rollout following.

Stripe's Agentic Commerce Protocol (ACP) is an open-source specification developed with OpenAI. The protocol defines how agents and businesses transact, including checkout coordination and secure credential sharing. It can be implemented as a RESTful interface or MCP server.

Stripe also built Shared Payment Tokens (SPTs), a new primitive that lets agents initiate payments using buyer permission without exposing credentials. SPTs are scoped and time-constrained to single transactions, work with Stripe's Radar fraud detection, and integrate directly with PaymentIntents.

One important limitation: ACP v1 only works with Stripe-connected merchants. If a merchant doesn't use Stripe as their payment processor, they cannot accept an SPT. This makes ACP less of an open standard and more of a Stripe ecosystem feature. For agents that need to transact across the broader merchant landscape, including non-Stripe merchants, alternative approaches like card rails or execution layer solutions remain necessary.

Google's A2A Protocol enables communication between AI agents across platforms. Originally launched with over 50 technology partners including PayPal, Salesforce, and SAP, the protocol is now donated to the Linux Foundation with support from over 150 organizations.

Google extended A2A with the Agent Payments Protocol (AP2), which provides payment-method-agnostic infrastructure for secure agent transactions. AP2 uses "mandates," cryptographically-signed digital contracts that provide verifiable proof of user intent.

The Universal Commerce Protocol (UCP) establishes a common language for commerce between consumer surfaces, businesses, and payment providers. Developed with Shopify, Etsy, Wayfair, Target, and Walmart, UCP is endorsed by Visa, Mastercard, Stripe, American Express, and others.

The Landscape at a Glance

| Layer | Key Players | Primitive | Status | |-------|-------------|-----------|--------| | Infrastructure | PayOS, Nekuda, Prava, Proxy | Agent Tokens, Mandates, Multi-agent Wallets, Dedicated Cards | Live | | Identity | Skyfire, Visa TAP, Agentic Commerce Consortium | KYA Protocol, Trust Scores, Cryptographic Identity | Live/Pilots | | Issuance | Lithic, Marqeta, Highnote, Rain | Programmable Cards, Authorization Stream, Stablecoin Settlement | Live | | Execution | Rye, Induced AI | Browser Automation, Fraud Mitigation | Live | | B2B | Natural, Payman, Brex, Ramp | Workflow Payments, Human-in-Loop, Expense Agents | Live/Beta | | Network Standards | Visa TAP, Mastercard Agent Pay, Stripe ACP, Google A2A/AP2 | Protocol Specifications | Live/Draft |

Where Proxy Fits

Our approach at Proxy centers on dedicated cards running on existing rails. Each agent or workflow gets its own virtual card with attestation-before-access: agents must prove authorization before any transaction attempt is processed.

This architecture makes specific tradeoffs. The overhead of managing individual cards is higher than shared tokens. But the isolation provides cleaner blast radius when something goes wrong and simpler audit trails for compliance.

We're not trying to replace card networks. We're wrapping them for agents, adding the programmatic controls and identity verification that autonomous systems require while maintaining universal merchant acceptance.

The dedicated card model works especially well for high-stakes use cases where spend isolation matters: regulated industries, financial services, enterprise procurement. For high-volume, low-value transactions, shared token approaches may prove more efficient.

What Comes Next

The current landscape is modular by necessity. Different companies have specialized in different layers, and integration requires stitching multiple vendors together.

This will consolidate. The eventual winners will likely offer identity, logic, payment rails, and controls in one stack. We're already seeing this with Skyfire combining identity and payments, Visa building out Intelligent Commerce as an end-to-end framework, and Stripe packaging ACP with their existing payment infrastructure.

Several dynamics will drive the next phase:

Merchant adoption of agentic protocols. The Agentic Commerce Consortium's work on merchant opt-in standards will determine how quickly retailers enable agent-friendly checkout. Visa predicts mainstream adoption by holiday 2026, but that depends on merchant infrastructure catching up.

Protocol standardization. Multiple competing standards from Visa, Mastercard, Stripe, and Google create integration complexity. Either one will win or an interoperability layer will emerge. The Linux Foundation's involvement with A2A suggests the industry wants open standards.

Trust score maturation. Agent identity systems are nascent. As transaction histories accumulate, something like agent credit scores will emerge, enabling dynamic trust levels and spending limits. Skyfire's KYA protocol is an early step in this direction.

Regulatory clarity. Who is liable when an agent makes a fraudulent purchase? How do consumer protection laws apply to autonomous transactions? Regulators are watching this space closely, and their guidance will shape architectural decisions.

The agent payments landscape in 2026 is genuinely new territory. A year ago, this infrastructure barely existed. Now there's a real ecosystem with funded companies, production transactions, and competing standards.

Whether you're building an AI agent platform, researching payment options for autonomous systems, or just trying to understand where the industry is headed, the fundamental reality is this: AI agents are becoming economic actors, and the infrastructure to support them is being built right now.

---

Proxy provides virtual cards and payment infrastructure for AI agents, with dedicated cards per workflow and attestation-before-access controls. Learn more about enabling programmatic payments for your autonomous systems.