An AI agent can buy from a million Shopify stores without a human clicking checkout. The legal framework is from 1986.

What Wildcard Actually Does — and Why YC Wrote a Check

Wildcard entered Y Combinator's Winter 2025 batch and raised a $500,000 seed round, a bet that the next generation of API infrastructure won't be built for developers at all, but for AI agents making purchases on their behalf.

Founded in 2025 by Kaushik Mahorker and Yagnya Patel, the company positions itself as the gateway layer that lets AI agents find, select, and execute API calls using natural language instead of hand-written tool definitions and prompt-tuned LLM chains. The core product is an open-source registry and SDK built around a format called agents.json — a structured contract that API providers publish so agents can discover and call their endpoints reliably. An agent queries Wildcard in natural language, gets back the right API flow, and executes it through the SDK on the developer's own infrastructure. At launch, Wildcard offered ten curated APIs with verified integrations from providers like Resend and Alpaca Markets. The company says developers can add a new integration in under 90 seconds.

That framing (outcomes over discrete actions) is what separates Wildcard from a standard API directory. Mahorker and Patel identified a problem: existing APIs expose low-level endpoints. An agent trying to buy NVIDIA stock, send an email, and post a Slack message has to chain multiple discrete calls with bespoke tool definitions for each one. Wildcard bundles those into a single agent-callable flow.

The company operates two tracks simultaneously. One is the agentic API gateway described above. The other is a more conventional (and already revenue-generating) analytics platform for e-commerce brands. Under the AEO/GEO (Answer Engine Optimization / Generative Engine Optimization) banner, Wildcard helps merchants track how their products appear across ChatGPT, Google AI Overviews, Gemini, and Amazon Rufus. It monitors product mentions, identifies visibility gaps, and uses AI agents to enrich product data, generate comparison pages, build FAQs, and improve off-site discoverability. The company says 67% of products lack the attributes AI engines need to recommend them, and that competitors average 43 more external mentions (numbers from its own platform analytics).

That e-commerce business is also where agentic commerce enters the picture directly. Wildcard's site advertises early access for brands that want to enable instant checkout inside ChatGPT and Gemini, supporting the Agentic Commerce Protocol (ACP) for ChatGPT and the Universal Commerce Protocol (UCP) for Gemini and Google AI Mode. This is the bridge between the two halves of the company: the same infrastructure that helps agents find and call APIs is the infrastructure that lets an AI shopping agent complete a purchase.

The filing and the founders. The dealroom.co filing confirms the round closed on March 12, 2025. The company is based in San Francisco and was still small at the time of its YC launch. The Y Combinator company page lists it as active with Mahorker as the listed founder, and the LinkedIn post from YC announcing the launch drew 1,219 reactions and 71 comments, suggesting genuine community interest.

Mahorker's background fits the problem. He was an Engineering Manager at Scale AI, where he led the GenAI Allocation team and built an e-commerce enrichment engine that processed 2.4 million attributes across 400,000 SKUs. Before that he worked at AWS Elastic File System. Patel's experience spans NLP and knowledge graphs at Tesla, Amazon, and Truveta. They met as undergrads at UCSB eight years ago.

The signal YC likely saw: Wildcard is not building another LLM wrapper. It is building the integration layer that sits between agents and the real-world APIs those agents need to act on — payments, inventory, shipping, communication. If agentic commerce scales, someone has to own that middleware. Wildcard wants to be the registry every agent checks first.

The OpenAI-Stripe Protocol That Makes Agentic Commerce Real

The reason Wildcard can exist at all — the reason any startup can build a gateway layer for AI agents executing real transactions — traces back to a single piece of infrastructure that didn't exist a year ago. In late 2025, Stripe and OpenAI co-developed the Agentic Commerce Protocol (ACP), an open standard that defines how AI agents, buyers, and businesses complete purchases together. It is the technical substrate that turns "AI agent commerce" from a demo into something that can process a payment.

Before ACP, the idea of an AI agent buying something on your behalf was mostly theoretical. An agent could browse, compare, recommend — but the moment money needed to move, you hit a wall. Every merchant integration was bespoke. Every agent platform needed its own custom checkout pipeline. Stripe's own research, published when they announced the protocol, put the problem plainly: businesses wanted to reach customers through agentic channels but didn't want to build and maintain dozens of one-off integrations. The AI economy needed standardized infrastructure or it would fragment into incompatible silos.

ACP is that standard. It is open source, Apache 2.0 licensed, and available for any business or AI platform to implement. The specification covers the full transaction lifecycle: agentic checkout (creating, updating, and completing checkout sessions), cart and product feed management, delegate payment through secure token passing, and order tracking via webhooks for confirmation, shipping, delivery, and refunds. It works with REST or MCP transport, connects to any commerce backend or payment processor, and is PCI compliant — payment credentials are passed as secure tokens, never exposed in the clear.

The protocol's first production deployment is Instant Checkout in ChatGPT, which lets US users buy directly from Etsy sellers and, coming soon, from over a million Shopify merchants including Glossier, Vuori, Spanx, and SKIMS — all without leaving the chat interface. Stripe powers the payment layer. OpenAI powers the agent surface. ACP is the connective tissue between them.

The transaction flow is straightforward. A buyer discovers a product through an AI agent, selects what they want, and grants permission to initiate checkout. The agent sends a checkout request to the business along with a secure payment credential token. The business (which remains the merchant of record) can accept or decline using its own fraud and risk signals. The payment provider processes the tokenized credential. At no point does the AI agent see or store raw card data.

What makes ACP structurally significant is what it avoids. The merchant keeps control over which products are sold, how they're presented, how transactions are processed, and how orders are fulfilled. Businesses don't need to build a new integration for every AI agent that comes along — they implement ACP once and distribute to any compatible agent. The protocol supports physical goods, digital goods, subscriptions, asynchronous purchases, and multi-merchant carts. Stripe says the spec is designed to eventually support custom checkout capabilities like in-store pickup and dynamic pricing through agents.

Meta has since joined as a technical steering committee member, and the GitHub repository (1.4k stars, 226 forks, 27 contributors as of the latest commit in June 2026) shows active development on feed versioning, order schema refinements, and fulfillment event modeling. The spec is still in beta. The latest stable release snapshot is dated 2026-04-17, with an unreleased development branch tracking changes for the next version.

For startups like Wildcard, ACP is the reason the category is buildable. Without a shared protocol, an API-gateway layer for agentic transactions would need to negotiate custom integrations with every merchant and every agent platform independently — a scaling problem that would kill most early-stage companies before they shipped. ACP collapses that problem into a single implementation target. Build to the spec, and you can theoretically route transactions between any compatible agent and any compatible merchant.

Stripe's documentation notes that if you already process payments with Stripe, enabling agentic payments can require changing as little as one line of code. That's the pitch: the hard infrastructure work is done. What's left is the product layer (the orchestration, the agent-tool integration, the transaction-state management) which is exactly where companies like Wildcard are placing their bets.

What Kind of Engineers Agentic Commerce Startups Actually Need

The job postings are specific enough to read like an accidental manifesto. Accenture's listing for a Technical Commerce & AI Consultant calls for someone who can build multi-agent architectures coordinating planners, retrievers, reasoning agents, and optimization components — then wire those systems into enterprise commerce platforms like Adobe Commerce, Salesforce Commerce, and SAP CX. The candidate needs Python, SQL, FastAPI, vector databases, and hands-on experience with LangChain, LlamaIndex, CrewAI, AutoGen, or LangGraph.

That listing describes a role that barely existed eighteen months ago. The skill stack is worth unpacking because it reveals what agentic commerce actually demands from engineers — and why it doesn't map onto any single existing job category.

API orchestration is the core competency. An agentic commerce system doesn't call one endpoint and return a result. It coordinates across product inventory APIs, pricing engines, payment processors, personalization services, and fulfillment systems — often mid-transaction, with the agent making real-time decisions about which tools to invoke and in what order. The Accenture listing calls this "schema-driven tool invocation," referencing standards like MCP (Model Context Protocol). Amazon's Principal TPM posting describes the output of this work as "robust, scalable APIs and integration layers that connect Amazon's services with third-party platforms." The person building these layers needs to think in terms of stateful, multi-step workflows where a failure at step six means rolling back steps one through five — or at minimum, logging exactly what went wrong and why.

Agent-tool integration is a distinct discipline from traditional backend engineering. A conventional integration engineer connects system A to system B with a defined data contract. An agent-tool integration engineer builds interfaces that an LLM-powered agent will discover, evaluate, and invoke autonomously — which means the tool descriptions, input schemas, error responses, and latency characteristics all become part of the agent's reasoning context. Wildcard's Founding Engineer listing on Y Combinator's board makes this explicit: the role involves "orchestrating browser-based scraping and retrieval infrastructure at scale across AI commerce surfaces" and "defining how our product adapts to emerging agentic commerce protocols and platform launches." The engineer isn't just building a pipeline. They're building a system that has to interpret and respond to protocols that are still being written.

Transaction-state management is where agentic commerce gets genuinely hard. When a human clicks "buy," the checkout flow is linear and the user is present to correct errors. When an AI agent initiates a purchase, the system has to handle authorization, payment capture, inventory holds, shipping confirmation, and potential refunds — potentially without a human in the loop for any individual step. JPMC's Vice President role in agentic commerce lists "integrating AI models with payment networks" and "ensuring compatibility with merchant systems and payment protocols" as core responsibilities. This is financial infrastructure work layered on top of AI orchestration work, and the intersection is where most of the unsolved engineering problems live.

The talent market is responding, if slowly. LinkedIn lists 278 agentic commerce jobs in the United States as of June 2026, with 19 new postings added in the most recent cycle. The roles are multiplying, the salary bands are competitive, and the skill requirements are converging around a profile that didn't have a name until recently.

The engineers who will do best in this space are the ones comfortable with ambiguity as a permanent condition. Wildcard's job description says it directly: "energized by ambiguity, speed, ownership, rapid pivots, and direct product decision-making." The protocols are shifting. The platforms are launching and changing. The agent behaviors are unpredictable in production. This is not a domain where you can spec a system, build it, and hand it off. It's a domain where the engineering is the product, and the product is a moving target.

Who Else Is Building in Agentic Commerce

Wildcard didn't appear in a vacuum. The $500K YC seed bet makes more sense when you see the ecosystem forming around it — a loose constellation of startups and incumbents all racing to own some layer of the transaction stack when the buyer is an AI agent instead of a human with a credit card.

Stripe is the most obvious signal. The company co-developed ACP with OpenAI, and its own hiring reflects the bet: Stripe added 39 roles to the Zero G Talent board in the past week alone, including a dedicated Program Manager, Agentic Commerce Go-to-Market spanning five U.S. cities. That's not a research experiment. That's a go-to-market hire, someone tasked with selling the infrastructure to merchants who need to accept agent-driven payments. When Stripe staffs a role like that, it's because the pipeline of merchants asking about agentic checkout is real enough to justify headcount.

OpenAI's presence is more indirect but no less significant. The company's 38 new listings on the board include forward-deployed engineers and research roles tied to proactive AI systems — the kind of agents that would initiate a purchase without a human clicking "buy." OpenAI isn't building a commerce product itself. It's building the agent intelligence that makes commerce products like Wildcard necessary.

Beyond the two protocol co-developers, the agentic commerce stack is filling in from multiple directions. On the discovery side, startups working on generative engine optimization — making products visible to AI agents the way SEO made them visible to search engines — are raising early rounds and hiring small integration teams. On the agent side, AI shopping assistants that can compare, negotiate, and transact on a user's behalf are pulling in seed funding, often from the same crypto and fintech investors who backed the last wave of checkout startups. On the infrastructure layer, companies building authorization frameworks, transaction-state managers, and fraud-detection systems tuned to non-human buyers are emerging as the plumbing that makes the whole stack work.

What ties these categories together is a shared technical problem: existing payment infrastructure assumes a human is on the other end of the session. Credit card flows, 3D Secure, even basic cart logic — all of it was designed around a person making a conscious decision at checkout. Agentic commerce breaks that assumption. An AI agent doesn't have a billing address in the traditional sense. It doesn't authenticate with a fingerprint. It may execute a transaction on behalf of a user who authorized the goal ("book the cheapest flight to Lisbon next Friday") but not the specific merchant or price. Every layer of the stack (discovery, selection, authorization, payment, confirmation) needs to be re-engineered for that reality.

Wildcard's bet is that the API-gateway layer is the right place to start. But the hiring signals from Stripe, the research direction at OpenAI, and the breadth of startups entering adjacent layers all point to the same conclusion: agentic commerce is not a one-company thesis. It's a category forming in real time, and the engineering roles it demands (API orchestration, agent-tool integration, transaction-state management) are already showing up on job boards. The question isn't whether this stack will get built. It's which layers will consolidate and which will stay fragmented long enough for startups to carve out defensible positions.

Why AI Engineers Should Pay Attention Now

The AI talent market in 2025 is a story of two simultaneous truths: near-universal adoption and acute scarcity. McKinsey's 2025 State of AI survey found that 88% of organizations now use AI in at least one function, up from 78% a year earlier. But the supply side hasn't kept pace. Rytsense Technologies reports that for every qualified AI candidate, 15 to 20 open positions compete. Universities produce roughly 40,000 AI graduates globally each year against demand for over 500,000 new roles. That gap isn't closing — it's widening.

What's changed in 2025 is the composition of that demand. For the past three years, the bulk of AI hiring flowed toward a relatively well-defined set of roles: LLM fine-tuning, RAG pipeline engineering, copilot integration, MLOps. Those roles aren't going away, but they're no longer where the sharpest demand signal is coming from. The new pull is from agentic systems — and agentic commerce in particular represents a distinct enough engineering challenge that it's quietly creating its own hiring category.

The demand signal is already visible in the data. Magnit Global's workforce data shows AI/Automation role fills doubled year-over-year in Q1 2025, jumping from 3% to 6% of total fills — even as overall IT/Tech fills contracted by 2%. More telling is the internal shift: within that AI/Automation category, automation roles grew from 32% to 44% of fills, while data engineering dropped from 46% to 32%. Organizations aren't just hiring more AI people. They're hiring different ones — people who can wire systems together, not just train models.

Agentic commerce sits at the intersection of several of those emerging specializations. Visa's recently unveiled Trusted Agent Protocol — developed in response to a 4,700% surge in AI-driven traffic to U.S. retail sites — is another signal that the infrastructure for agentic commerce is being built now, and it needs engineers who understand both the agent side and the transaction side. McKinsey's agentics data reinforces the enterprise demand side: 23% of organizations are scaling AI agents in at least one function, with IT, knowledge management, and engineering leading. But the Futurum report adds a critical caveat — 60% of DIY agentic AI initiatives fail to scale past pilot stages, largely due to governance gaps and integration complexity. That failure rate is, paradoxically, a hiring signal. Enterprises that can't build agentic systems in-house will buy them, and the companies building them need engineers who've already solved the problems that kill pilots: authorization flows, error recovery in multi-step transactions, observability across agent tool calls.

This matters for talent allocation because agentic commerce engineering doesn't map cleanly onto existing role definitions. A senior machine learning engineer who can fine-tune a 70B-parameter model may have no experience with idempotency keys, webhook verification, or payment state machines. An API integration engineer who's spent years on REST endpoints may not understand tool-use protocols or agent planning loops. The engineers who can bridge that gap — who understand enough about how agents make decisions to build reliable transaction infrastructure around those decisions — are scarce. They're also expensive.

Wildcard's $500K YC bet is small in absolute terms, but it's a directional indicator. The company is hiring for the exact profile that's becoming the bottleneck in agentic commerce: engineers who can build API-gateway layers that let AI agents execute real transactions reliably. Zero G Talent's board shows OpenAI adding 38 roles in the past week alone, many in research engineering and forward deployment — roles that increasingly overlap with the agentic systems Wildcard is building around. Stripe's 39 new listings span billing infrastructure and agentic commerce go-to-market. These aren't coincidences. They're the talent market repricing around a new category of work.

For AI engineers evaluating where to allocate their careers in 2025, the calculus is straightforward: LLM training and RAG pipeline work are becoming commoditized. The scarce, defensible skill is building the infrastructure that lets agents act in the real world — especially where money changes hands. Wildcard is one of the first pure plays in that space. It won't be the last.

The Regulatory and Trust Layer Nobody Is Talking About

The OpenAI-Stripe protocol gave AI agents the ability to transact. It did not answer the harder question of what happens when one of those transactions goes wrong — or when an agent buys something its human principal never intended.

That gap is where the next wave of engineering hiring in agentic commerce will happen, and it has almost nothing to do with APIs.

The authorization problem is unsolved. When a user gives an AI agent credentials to act on their behalf, the legal meaning of that authorization is genuinely unclear. Amazon's November 2025 lawsuit against Perplexity AI put this in stark relief: Amazon alleged that Perplexity's Comet browser accessed Amazon accounts and transmitted data without authorization, violating the Computer Fraud and Abuse Act — a statute written in 1986, before the concept of an autonomous purchasing agent existed. In March 2026, Judge Maxine M. Chesney of the Northern District of California granted a preliminary injunction finding "strong evidence" Perplexity continued accessing Amazon's systems after a cease-and-desist. The Ninth Circuit heard oral argument in June 2026. One judge on the panel, District Judge John Hinderaker, said plainly: "This case is difficult in part because we are dealing with a statute from 1986. It's not really built for these circumstances." The panel has not yet ruled.

The case exposes a doctrinal vacuum. Existing frameworks (agency law, product liability, the EU's AI Liability Directive) each cover fragments of the problem. None cleanly addresses a scenario where an AI agent with valid credentials makes a purchase that exceeds its user's intent, or where a platform revokes access but the agent persists. A report from the American Enterprise Institute's AI Liability Project taxonomizes three policy approaches: traditional tort liability, conditional immunity in exchange for third-party auditing, and no-fault compensation schemes. None has been adopted for agentic commerce specifically.

Federal procurement policy is ahead of consumer law — barely. OMB Memorandum M-25-22, issued April 3, 2025, requires federal agencies acquiring AI systems to convene cross-functional teams, test proposed solutions before award, and include contractual terms for ongoing performance monitoring, data portability, and vendor lock-in protections. Agencies must update internal acquisition procedures by December 29, 2025. For "high-impact" AI use cases — those affecting civil rights, access to government services, or critical infrastructure — M-25-21 requires pre-deployment testing, AI impact assessments, and human oversight mechanisms. The General Services Administration is building a web-based acquisition resource repository due October 20, 2025.

These are procurement safeguards, not consumer protection rules. But they establish a vocabulary (impact assessments, cross-functional review, performance-based contracting) that agentic commerce startups will need to adopt if they want to sell into the enterprise or government market.

The trust engineering gap is the hiring frontier. The roles that don't yet have job titles are the ones that matter most. Agentic commerce needs engineers who can build transaction-state systems that log not just what an agent bought, but the chain of reasoning and authorization that led to the purchase. It needs fraud-detection models that distinguish between a user changing their mind and an agent behaving outside its scope. It needs credential-management systems that can enforce granular spending limits, category restrictions, and real-time revocation — without breaking the speed that makes agentic commerce useful.

McKinsey's analysis of the agentic commerce opportunity notes that identity verification, buyer protection, and fraud detection become "core components" of the payments layer, not afterthoughts. Crowe LLP's risk management team frames agentic commerce as a "regulatory stress test" for AML programs and consumer protection laws designed for human-initiated transactions.

What to build toward. The Ninth Circuit's eventual opinion in Amazon v. Perplexity will set the first real precedent for how legacy access statutes apply to autonomous agents. Until then, every team building agentic integrations is operating in a legal gray zone — one where the engineering decisions made today about credential handling, logging, and authorization scoping will determine liability exposure tomorrow.

For engineers watching this space, the signal is clear: the hard problems in agentic commerce are no longer about making agents transact. They're about making them transact responsibly, accountably, and defensively. That's a different job description entirely.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at OpenAI and Stripe, and the people building the field.