Software Is Becoming Labor: The Structural Shift Behind YC Spring 2026
For most of the last decade, the most coveted seat in a YC batch was the one occupied by a founder building a new AI model or a smarter chatbot. In Spring 2026, the coveted seat shifted. The 196-company cohort — 61% B2B, with a combined 91% across enterprise, industrials, fintech, and healthcare — was dominated not by AI agents themselves but by the infrastructure layer underneath them.
The batch’s defining thesis, as observed by investors who attended Demo Day, was blunt: “software is becoming labor.” That phrase captured something more precise than a marketing slogan. These founders are not building tools that help humans work faster. They are building systems that perform work autonomously — and more importantly, they are building the rails, wallets, memory stores, and trust layers that those autonomous systems will need to operate at scale.
The distinction matters commercially. An AI agent that can research, decide, and execute is only as powerful as the infrastructure it can call on. If it cannot pay for something without a human approving each transaction, it is not autonomous. If it cannot prove its identity to a merchant API, it bounces. If it has no persistent memory, it starts every task from scratch. The YC Spring 2026 batch recognized this gap and flooded into it.
The Picks and Shovels Playbook: Five Infrastructure Layers in Demand
Investors and analysts tracking the batch have identified at least five distinct infrastructure layers attracting significant startup attention. Each one corresponds to a friction point that currently prevents AI agents from operating independently in commercial environments.
1. Payments and Card Issuance: Giving Agents a Wallet
The most funded sub-category in agentic commerce infrastructure is payments. Basis Theory, a card vault and tokenization startup, closed a $33 million Series B in October 2025 led by Costanoa Ventures, bringing its total raised to approximately $50 million. Nekuda, which provides AI-native wallets using tokenized card infrastructure, raised a $5 million seed round in May 2026 backed by Madrona, Amex Ventures, and Visa Ventures. Skyfire, focused on micropayment rails for agents, raised $9.5 million across seed and follow-on rounds, with backing from Neuberger Berman and a16z CSX.
The logic is straightforward: agents need their own financial identity. Virtual card issuance platforms like Lithic, Stripe Issuing, and Marqeta are all seeing increased interest from agent-first startups that need programmatic spend controls — the ability to issue an agent a card with hard limits, merchant-category restrictions, and real-time kill switches. Without this layer, every agent purchase requires a human to approve a charge, which negates the autonomy that makes agents valuable in the first place.
According to data from Visa, agent traffic to retail and merchant sites grew 1,200% year-over-year in 2025 — a signal that the commercial transaction volume is already arriving even as the infrastructure to handle it cleanly is still being built.
2. Identity and Authorization: Proving the Agent Is Who It Claims to Be
Payments without identity verification create fraud. Agentic commerce without robust identity infrastructure creates a compliance nightmare. This is the second major infrastructure layer the YC batch is targeting.
Mastercard is developing what it calls “verifiable intent” — tamper-resistant authorization records that create a traceable audit trail for agent-initiated payments. American Express has launched Agent Purchase Protection for registered agent transactions. These moves from the incumbents signal that the identity problem is real and urgent, and they create a market for startups that can work within — or across — existing rails.
Startups in this space are building agent credential systems, scoped authorization tokens (similar to OAuth but designed for machines rather than humans), and behavioral fingerprinting that can distinguish a legitimate AI agent from a fraudulent automated script. Forter, tracking fraud patterns in this space, observed an 18,510% day-over-day increase in agent traffic following ChatGPT’s commerce launch, alongside a 50% rise in fraud using scripted and automated attack modes. The identity infrastructure layer is not a nice-to-have; it is a prerequisite for the agent economy to scale without becoming a fraud vector.
3. Memory and Context: Making Agents Actually Useful Across Sessions
An agent that forgets every preference, every past purchase, and every workflow context after each session is an agent that is frustrating rather than useful. Memory and context layers are the third major infrastructure category attracting YC-adjacent founders.
These systems range from simple persistent key-value stores to sophisticated “company brain” products that index an organization’s internal knowledge — past contracts, supplier relationships, approval workflows — and make it queryable by autonomous agents in real time. The YC batch included multiple companies in this cluster, building memory runtimes, vector search layers optimized for agent retrieval, and context APIs that allow agents to pick up a task precisely where they left off.
The commercial model here is typically infrastructure-as-a-service: per-query pricing or per-seat licensing to enterprise customers deploying internal agents. The defensibility comes from data flywheel effects — the longer an enterprise uses a memory layer, the richer and more valuable that context store becomes, creating switching costs that purely workflow-level automation tools lack.
4. Monitoring and Observability: Knowing When Agents Fail
No enterprise will deploy autonomous agents at scale without visibility into what those agents are doing. Agent observability — logging decisions, flagging anomalies, tracking task completion rates — is the fourth infrastructure layer the batch is addressing.
This is analogous to the DevOps observability stack that emerged around microservices: Datadog, New Relic, and their successors became massive businesses not by building services themselves but by instrumenting everyone else’s services. YC Spring 2026 companies in this cluster are building agent-native equivalents: dashboards that show not “is this API responding” but “is this agent making decisions consistent with its instructions” and “is it staying within its authorized scope.”
The risk profile for enterprises is asymmetric: a rogue agent that exceeds its authorization or makes a wrong purchase on a commercial account creates legal, financial, and reputational exposure. Observability tools that catch these events and route them for human review are a compliance purchase as much as a performance tool.
5. Checkout Execution and Merchant Protocols: Closing the Last Mile
The final infrastructure layer is perhaps the most tangible: the actual mechanics of an agent completing a purchase on a website or marketplace that was not designed to be transacted with by a machine. Adobe conversion data shows that AI-referred storefronts convert at 31% higher rates and generate 254% more revenue per visit than traditional traffic — which means the demand for agent-driven commerce is already commercially significant.
Startups like Rye (universal checkout API requiring no merchant integration), Zinc (purchasing API covering Amazon, Walmart, Target, and Best Buy), and Firmly (which raised $5.2 million to aggregate across protocols including ACP, UCP, and AP2) are building the plumbing that allows an agent to execute a transaction at the checkout layer without requiring every merchant to build bespoke integrations. The fragmentation problem is real: no single integration currently covers the full open web, and long-tail merchants remain largely invisible to agent purchasing systems.
Advertisement
What This Means for Founders, Investors, and Enterprise Teams
YC Spring 2026’s infrastructure tilt is not accidental — it reflects a maturing understanding of where durable value accretes in a new technology wave. The agent layer itself may commoditize quickly. The infrastructure underneath it, if positioned correctly, will not.
1. Founders: Own a Workflow, Not Just a Tool
The sharpest investor critique of the Spring 2026 batch, as noted by observers at Demo Day, was that some infrastructure companies were “layers that can be useful without becoming durable businesses.” The distinction between useful and durable comes down to workflow ownership. A memory layer that integrates deeply into an enterprise’s approval workflows, supplier databases, and compliance rules is sticky. A generic vector store is not. Founders building infrastructure need to pick a vertical, get embedded in the operational reality of that vertical, and own outcomes — not just outputs.
2. Investors: The Durable Money Is Below the Agent Layer
Q1 2026 saw $242 billion flow to AI companies globally, representing 80% of all venture funding — the highest proportion on record. But concentration is extreme: four companies (OpenAI, Anthropic, xAI, Waymo) captured 65% of that total, raising a combined $188 billion. For investors who cannot write $10 billion checks, the infrastructure layer beneath the frontier model giants represents the more accessible — and potentially more defensible — opportunity set. Early-stage deal count fell 30% at seed even as seed dollar volume rose 31%, suggesting capital is concentrating in fewer, larger infrastructure bets rather than spreading across many thin agent plays.
3. Enterprise Teams: Procurement Must Account for the Agent Stack
Enterprises evaluating AI agent deployments in 2026 need to think beyond the agent vendor and assess the full stack: which payment rails will agents use, how will agent identities be credentialed and scoped, what memory layer will persist context across sessions, and who is responsible when an agent exceeds its authorized spend. These are procurement questions, not just technical ones, and they need to be answered before autonomous agents touch commercial workflows at scale.
The Correction Scenario: When Infrastructure Bets Go Wrong
Not every infrastructure play will compound. The history of “picks and shovels” investing carries cautionary tales alongside the success stories. The dominant risk in the agent infrastructure category is timing: building a specialized agent payment rail six months before Stripe ships a native agent wallet product means competing against a well-capitalized incumbent with existing merchant relationships.
The second risk is fragmentation lock-in going the other way — if the major AI platforms (OpenAI, Google, Anthropic) build proprietary infrastructure for their own agents, third-party infrastructure startups may find their addressable market shrinking to niche use cases. This is the scenario Mastercard CEO Michael Miebach acknowledged when describing the company’s engagement with Google, Microsoft, and OpenAI to “shape what comes next” — the incumbents are not passive bystanders in the infrastructure race.
The startups that will survive this correction scenario are the ones that become so embedded in specific regulated verticals — healthcare billing, insurance underwriting, government procurement — that a generic platform alternative cannot easily displace them. ChatGPT already fields 50 million shopping-related queries daily, and the agentic commerce market is projected to grow from a $135 billion TAM in 2025 to $1.7 trillion by 2030 at a 67% CAGR. The infrastructure layer to support that volume does not yet fully exist. The Spring 2026 YC batch is the first organized wave of founders trying to build it — and the ones who survive will not be the ones who built the most elegant layer, but the ones who went deepest into the messiest operational problems.
Frequently Asked Questions
What is the “picks and shovels” strategy in the context of AI agents?
The picks and shovels strategy refers to investing in or building the enabling infrastructure that AI agents need to function — payments, identity, memory, and observability tools — rather than building the agents themselves. Just as gold rush miners needed pickaxes regardless of whether they struck gold, AI agents will need reliable infrastructure regardless of which specific agent products dominate the market. This infrastructure layer is often more defensible and less subject to rapid commoditization than the agent layer.
Why are payments and identity infrastructure the most funded categories in agentic commerce?
Payments and identity are the most funded because they represent the hardest blockers to autonomous agent operation. An agent that cannot transact independently or prove its identity to a merchant API cannot complete commercial tasks without constant human approval, which undermines the core value proposition. Companies like Basis Theory ($33 million Series B), Nekuda ($5 million seed), and Skyfire ($9.5 million) attracted capital precisely because they address this bottleneck directly. Payment networks Visa, Mastercard, and American Express are also building in this space, validating the market’s importance.
How should enterprise teams evaluate AI agent vendors in 2026 given this infrastructure landscape?
Enterprise teams should evaluate AI agent vendors on five dimensions beyond the agent’s capabilities: (1) which payment rails the agent can access and with what spending controls, (2) how agent identity is credentialed and scoped to prevent unauthorized actions, (3) what memory layer persists context across sessions, (4) what observability tools exist to audit agent decisions, and (5) what the vendor’s roadmap is for regulated verticals like finance or healthcare. A capable agent built on weak infrastructure creates compliance exposure; procurement teams should treat the infrastructure stack as a first-class evaluation criterion alongside agent performance.
