Twenty-two companies were admitted to Mastercard Start Path's first Agentic Commerce & Services cohort, announced September 2–3, 2026 — the first time the program has assembled a full class around agent-initiated payments rather than folding agent tooling into its existing acceptance, blockchain, or security tracks. The Paypers reported on September 2 that "Mastercard's Start Path has admitted 22 companies to its first Agentic Commerce & Services cohort," spanning payments and AI infrastructure. Forkast's Tessa Vaughn framed the admission the following day as a signal "that the financial services establishment is ready to treat autonomous AI agents as legitimate participants in the payments ecosystem."

The timing is the part worth reading closely. Two days earlier, on August 31, Flagship Advisory Partners published a market tracker showing that across the same 80 large US and EU merchants it reviewed in December/January, "we have not seen any additional brands declare support for payments by autonomous AI agents." Zero movement at the demand end. Meanwhile one of the two global card networks just seeded twenty-two companies at the supply end — identity, rails, merchant enablement, fraud control, reconciliation, and liability insurance. Mastercard is not betting that merchants have adopted agentic payments. It is betting that the infrastructure to make them adoptable does not exist yet, and that whoever assembles it will route volume across its network. That is a different bet, with a different clock on it.

What Start Path Is, and What Changed Between January and September

Start Path is Mastercard's global startup engagement program, launched in 2014. Per Mastercard's January release, it has grown to include more than 500 companies from 60 countries. As of last September, the program had brokered 15,000 global connections, and participating startups had raised over $25 billion in capital after their time in the program. Simetrik's own announcement repeats those figures: more than 500 companies across over 60 countries, an estimated $25 billion in post-program capital, more than 15,000 connections between startups and Mastercard customers and partners worldwide.

The arc to this week began on January 20, 2026, when PYMNTS reported that Mastercard was expanding the program to include agentic commerce. That new track would focus, per the release, on "agent platforms and agentic enablement, agent-native tools and services, and emerging agentic payment flows." It joined existing Start Path programs focused on acceptance, blockchain and digital assets, emerging FinTech, open finance, small business, and security solutions. Applications opened that Tuesday.

That was the announcement of a door. This week is the announcement of who walked through it. The distinction matters because seven months elapsed between the two, and the intervening period is when Mastercard articulated what it thought it was buying. From the January release: "Intelligent agents can accelerate the pace of commerce only if the foundation beneath them is predictable, secure and universally understood. That's why we're building the standards, partnerships and capabilities that allow innovation to move at full speed without losing the coordination that keeps the system flowing." The expansion, the company said, would let it "engage with innovative newcomers who can bring the cutting-edge technology, products and services that make agentic commerce essential for people and businesses." Background from the same release — Mastercard joining Google's Universal Commerce Protocol, and work with Microsoft to bring Mastercard Agent Pay to Copilot Checkout — sets the direction, but neither is this week's news.

The Start Path Agentic Commerce & Services track is a dedicated global program. Its stated selection bar: startups typically have raised Seed or Series A+ funding, have a live product generating revenue, demonstrate clear differentiation and strategic alignment with Mastercard, and feature strong founding teams. The benefits are the interesting part for anyone modeling what a cohort seat is actually worth — access to Mastercard's global network of banks, merchants, and partners; dedicated support and tailored pipelines; customer and channel introductions; potential strategic investment via Mastercard's fund; community events and demo days, delivered in a virtual engagement format.

Vaughn's read on the strategy is the sharpest one published this week: "Mastercard's approach here is ecosystem-building rather than protocol-building. Rather than developing its own agent commerce standard, the company is using Start Path to identify and accelerate the startups that are building the infrastructure layer."

The 22, Grouped by Layer

Read as a list, the cohort looks scattered. Read as a stack, it is unusually legible — the selection covers roughly six functional layers, and each has more than one occupant.

Identity and agent credentials. Skyfire provides agents with verified identity and payment credentials across the open internet. t54 enables merchants and payment networks to safely transact with AI agents via real-time identity verification, risk assessment, and dispute resolution. Crossmint provides the wallet and credential infrastructure — described on the Start Path page as enabling fintechs to launch stablecoin products including wallets, on/off-ramps, orchestration, and embedded infrastructure. Vaughn identifies Crossmint and SolvaPay as representing "the two foundational layers of agent commerce: the payment rail and the identity layer."

Rails and settlement. SolvaPay lets businesses sell to and get paid by AI agents and their users. Nevermined operates as a commerce logic layer between AI agents and money — verifying, accepting, and settling agent-driven purchases. Merge helps fintechs and financial institutions securely connect products to business systems, agent tools, and AI models.

Merchant enablement. This is the largest group, which tracks with the Flagship finding that the merchant side is where the stall is. Firmly enables merchants to sell directly through AI agents and digital channels with zero engineering effort, offering instant checkout from discovery. Satsuma helps merchants launch and track LLM connectors without multiple bespoke integrations. ReFiBuy is an agentic commerce optimization platform for brands and retailers, structuring product catalogs for AI discovery, recommendation, and purchase. Grovity helps merchants align systems and operations with AI employees — an autonomous digital workforce. Thyris.AI builds agentic commerce and AI products for merchants and financial institutions across infrastructure, personalization, discovery, and autonomous experiences. Wizard is a shopping agent for ecommerce with personalized recommendations and a cross-retailer cart. 10Clouds helps enterprises deploy AI agents to automate complex processes and power agentic commerce. Sapiom is a platform that vertical AI companies run their agents on. Luca AI is an operating system for small businesses unifying sales, marketing, finance, and ops in a conversational interface with access to capital.

Fraud and control. Tunic Pay helps banks stop scams in agentic payments by assessing broader transaction context. PayOS is a payment intelligence layer providing details on every entity — card, account, business, person, transaction, agent — before money moves. Masraff unifies expenses, travel, and corporate cards with automated fraud prevention. Glassbox Labs builds enterprise-ready AI solutions for auditable workflows and production-grade payment experiences.

Reconciliation and audit. Simetrik provides AI-powered financial control infrastructure for agentic commerce, enabling transaction-level, audit-grade reconciliation.

Liability. Testudo offers AI liability insurance for U.S. companies managing the risks of generative AI deployment. MediConCen, adjacent, transforms insurance claims using agentic AI and blockchain.

The presence of a reconciliation company and an insurance company in a payments accelerator cohort is the tell. Networks do not add audit trails and liability cover to a stack they expect to stay experimental.

The Proof Points Already in Production

Three cohort members have shipped work with numbers or regulated structure attached, which distinguishes them from pitch-deck agentic commerce.

Simetrik disclosed in its September 3 release from San Francisco that more than 180 companies across over 50 countries use the platform to process 2.5 billion records every day. This is its second Start Path selection — the company notes it joined the Corporate Solutions track months earlier, and that "the new cohort is the first Start Path program focused specifically on Agentic Commerce and Services." In August it launched Simetrik Agent, described as "an autonomous agent that automates financial reconciliation and control tasks while operating on a deterministic, auditable core designed to keep people in control of critical decisions." The company's framing of the problem is the cleanest statement of why reconciliation belongs in this cohort at all: "As AI agents move beyond recommendations and begin executing transactions, finance teams will need a reliable way to track what happened across increasingly autonomous workflows."

Skyfire markets itself as "The Agent Trust Stack." Per the company, it provides AI agents with verified identity and payment credentials through its KYA protocol, enabling them to access websites, sign up, log in, and complete transactions across the open internet. It supports tokenized cards from networks including Visa, Mastercard, and Discover, as well as USDC stablecoins via its Agentic Wallet, which authorizes payments based on user mandates for specific merchants, products, and amounts. Mandate-scoped authorization is the mechanism that turns "an agent spent money" into "an agent spent money it was permitted to spend, on a thing it was permitted to buy."

Testudo is the one that prices the residual risk. It operates as a managing general agent and a Lloyd's of London coverholder, offering standalone generative AI liability insurance for U.S. companies. Its policies address coverage gaps created by new commercial general liability exclusions that took effect in January 2026 — meaning the standard corporate policy has been carved back specifically around AI. Testudo writes claims-made protection for hallucinations, IP infringement, unauthorized data disclosure, bodily injury, property damage, and regulatory violations, on A+ rated Lloyd's capacity, with limits reaching up to $9.25 million per insured (Fintech Global, March 9, 2026).

t54, which posted publicly about its selection, brings Know Your Agent verification and full traceability to the merchant/network side of the same trust problem Skyfire solves from the agent side.

The Week's Other Numbers: Supply Is Moving, Demand Is Not

Flagship's August 31 tracker, authored by Ben Brown, Katharine Watson, Alessandro Mighetto, and Pavle Stamenic, splits the market into three stages: Generative AI (customer-facing AI that generates or summarizes content), Agentic Commerce (agents that orchestrate multi-step shopping tasks but do not execute payment), and Agentic Payments (an agent initiates and completes checkout on the customer's behalf using permissioned tokens, virtual cards, wallets, or other rails).

The headline finding: "Only 4 of 80 companies support agent-led checkout, all of whom are US based." Salesforce became the third US software provider to support agentic payments, with AP2 and ACP support in what is now Agentforce Commerce. B2B software leads the ladder — Snowflake and IBM announcements pushed US brands to 7 of the top 10 supporting agentic commerce, and Amadeus, Experian, and Adyen took the EU count from 1 to 4. Consumer verticals plateaued: no new US entertainment and travel, retail, or food and drugs brands have announced since January.

The traffic side is not flat at all. Flagship cites Adobe data showing AI-driven traffic to travel and retail merchants reached new highs in Q2 — "but to be fair, it's still less than 1% of total retail visits" — with AI-referred visitors staying longer, bouncing less, and spending more. AI-driven traffic is up roughly 13x for retail and 20x for travel, and for retail those visits "generate 53% more revenue, a big shift from last year when AI-referred site visits were only worth half as much as visits from other sources." Walmart said items sold via its Sparky agent are up 4x year-over-year, with auto-reorder, recipe-based cart building, and meal planning. Amazon's Buy for Me, piloted in April 2025, completes purchases from third-party brand sites inside the Amazon app; Google's Universal Shopping Cart can buy from merchants including Wayfair, Chewy, and Shopify when a tracked price is hit.

Discovery works. Checkout does not. Flagship also scanned 20 large B2B tools — ERP and cloud accounting, spend management, AR/AP automation — and found "almost everyone has built and shipped native AI agents or MCP servers, but almost none of them will let an agent move money," with Ramp, Coupa, Melio, and BILL leading. And there is a cautionary data point in the record: Walmart's quick withdrawal from ChatGPT Instant Checkout in March, after finding it converted at only one-third the rate of normal website visits, "caused some retailers to step back from active projects on agentic payments."

For governance context: Mastercard was among the 26 founding members of Rain's Agentic Payments Alliance, launched August 18 alongside Visa, Fiserv, Circle, Solana, Remitly, and Avalanche — a working coalition whose members will set the charter together, with nothing ratified yet. Crossmint and PayOS sit in both APA and this Start Path cohort, which makes the layer distinction concrete: x402 is a wire format, APA is authorization, identity, fraud, loyalty, and advocacy, and Start Path is a builder pipeline into Mastercard's merchant and bank network. Governance one week, builders the next.

The Gap the Cohort Is Trying to Close

Vaughn states the technical problem precisely: "Most enterprise agents today can analyze data and make recommendations, but they cannot independently initiate financial transactions. The Mastercard cohort is designed to close that gap."

But the gap Flagship documents is not purely technical. Salesforce, Snowflake, IBM, Adyen and Experian shipped agentic capability; the top consumer brands did not follow, and Flagship's authors admit surprise: "In January, we forecasted that near-term progress is likely to focus on agent-assisted commerce... while agentic payments required more work to be done on trust, governance, and interoperability. The last six months have proved that to be true, though we are surprised by seeing zero movement among the top consumer brands for agentic payments."

Zero movement, against a Walmart conversion result that gave retailers a reason to pause, is a demand problem no accelerator fixes directly. What a cohort like this can do is remove the specific objections — a merchant cannot integrate (Firmly, Satsuma, ReFiBuy), cannot verify who is buying (Skyfire, t54, PayOS), cannot detect scams in agent traffic (Tunic Pay), cannot close the books afterward (Simetrik), and cannot get the general counsel comfortable with the exposure (Testudo). Each of the four brakes Flagship identifies as trust, governance, interoperability, and conversion has at least one company in this cohort pointed at it. Whether twenty-two seed and Series A companies can move eighty enterprise buyers inside a holiday cycle is the open question, not a resolved one.

The investor frame, via Simply Wall St's September 3 note, is blunter. With a market cap of about $515.2b, Mastercard "is using this agent focused Start Path cohort to explore how its existing payment rails could support AI driven transactions." If it works, "that keeps Mastercard at the center of payment initiation and authorization even when no human is clicking 'pay.'" The proposed scorecard: "how many Start Path cohort companies move into scaled commercial deployments over the next 12 to 24 months, measured by the volume of agent initiated transactions they route over Mastercard's network."

What to Do With This

Treat the cohort as a published map of the missing layers, not as an endorsement of twenty-two companies. Mastercard's selection criteria required live product and revenue, so each name marks a hole someone is already paying to fill. If your architecture has no answer at the identity layer, the fraud layer, or the reconciliation layer, the cohort tells you what a network thinks the answer should look like.

Assume the checkout bottleneck is commercial, not technical. Flagship's 4-of-80 number sits alongside 13x and 20x traffic growth and 53% higher retail revenue per AI-referred visit. Demand for agent discovery is real and measured; the block is at authorization, dispute, and audit — which is exactly where Skyfire's mandate-scoped KYA credentials, t54's Know Your Agent verification, PayOS's pre-money entity checks, and Simetrik's audit-grade reconciliation are aimed.

Price the liability explicitly. Commercial general liability exclusions for generative AI took effect in January 2026. Testudo's presence in a payments cohort, with Lloyd's capacity up to $9.25 million per insured, means the insurance market has already decided agent errors are a distinct risk class. If your agentic roadmap has no line item for it, your counsel will find one for you.

Watch three dates. Holiday 2026 remains the commercial deadline merchants are racing toward, and Flagship's read says consumer brands will not clear agentic payments in time. The CLARITY Act faces a Senate procedural vote on September 15, 2026 — nothing passed, but the crypto market structure outcome bears directly on the stablecoin rails Crossmint and Skyfire depend on. And Flagship says it will be watching B2B, travel, and food sectors closely in the coming months, which is where Ramp, Coupa, Melio, and BILL already have momentum.

Then judge the cohort on deployments, not press. Applications opened in January; the class was admitted this week; Start Path is recurring. The signal is which of these 22 reach Mastercard merchant and bank introductions and convert demo-day access into production volume over 12 to 24 months.

Map your stack against the six layers in this cohort — identity, rails, merchant enablement, fraud, reconciliation, insurance — and pick the hole to fill this week.

Sources