The Great Unbundling: When AI Agents Decide Where You Shop

Something fundamental shifted this week in how AI companies think about the physical world. While most of the industry obsesses over which model scores highest on benchmarks, the real action is happening at the borders — where AI agents try to actually do things, and the existing internet either welcomes them or slams the door.

Walmart just opened its door. Amazon bolted it shut. And in that divergence lies the most important strategic question facing every commerce company, SaaS platform, and data provider for the next decade: do you let AI agents in, or do you make them route around you?

1. Walmart Opens E-Commerce to AI Shopping Agents

Starting next month, Walmart will allow AI agents to browse, select, and purchase products on behalf of users — with approval controls and spending limits in place. The retailer published explicit terms of service for agents: they can't change payment methods, can't exceed the shopper's stated intent, and can only access information the customer has explicitly shared.

This stands in direct opposition to Amazon, which has spent the past month blocking Perplexity's shopping agent and Meta's Muse assistant from browsing its catalog. Where Amazon sees existential risk to its recommendation algorithm and Prime subscription moat, Walmart sees distribution.

The mechanics matter here. Walmart isn't just tolerating agents — it's building infrastructure for them. That means structured data feeds, programmatic access, and probably an agent-optimized pricing API within six months. They're treating AI agents like they treated mobile apps in 2010: as a new interface layer that will drive incremental volume.

🔥 Spark's Hot Take: Walmart is making the same bet that made them billions in the 1990s — be the low-price leader and let everyone else worry about margin compression. AI agents optimize for price and availability, not brand loyalty or algorithmic suggestions. Amazon's recommendation engine becomes worthless the moment an AI agent comparison-shops in real-time across retailers. Walmart just turned Amazon's greatest strength into a liability by being the first major retailer to say "yes, come in, buy stuff." Within 18 months, "AI agent checkout optimization" will be a standard e-commerce KPI, and every retailer that blocked agents will be scrambling to build integrations. The company that lost the internet to Amazon might win the agent economy by simply not fighting it.

2. Meta Creates Non-Voting Stock Class for AI Fundraising

Meta announced a new Class C non-voting share structure explicitly designed to fund AI investments while preserving Mark Zuckerberg's voting control. CFO Susan Li cited "significant" capital requirements for "infrastructure, potential acquisitions, and strategic initiatives" in the AI transformation.

The subtext is extraordinary. According to reporting, Meta's board feared that California's proposed billionaires' wealth tax could force Zuckerberg to relocate, disrupting the company's AI strategy. The solution: create a new share class that lets Meta raise hundreds of billions without diluting Zuckerberg's control by a single vote.

This isn't financial engineering — it's a declaration that AI capital requirements have blown past the scale where normal corporate governance can function. Meta needs to spend more than it can generate internally, but Zuckerberg trusts only Zuckerberg to allocate those dollars during the most consequential technology transition since mobile.

Here's what nobody's saying: institutional investors are about to put a number on exactly how much they'll discount a company where one person controls AI strategy indefinitely. That discount will become the reference price for every founder considering dual-class structures. If Meta's multiple compresses 10-15% relative to peers post-announcement, that's the cost of permanent control.

For founders, the lesson is simpler: if you think you might need to raise nine figures to compete in AI infrastructure, either structure for permanent control now or accept that you'll lose it when the capital requirements exceed your leverage.

3. PHP Security Holes Discovered by AI, Then Immediately Exploited

OpenAI and Google's Wiz used AI models to discover two remote code execution vulnerabilities in PHP, the programming language that powers 77% of websites. The vulnerabilities score 9.2 out of 10 on severity. The coordinated disclosure process, designed for human-speed patching, collapsed when the CVE IDs leaked a day early on forums. Exploits started circulating within hours.

A security firm that had originally reported one of the issues accidentally published a proof-of-concept video showing successful attacks on live targets, then deleted it — but copies spread across social media. PHP was forced to release patches 24 hours ahead of schedule.

This is the "defender's window" closing in real time. For months, OpenAI has argued that AI gives defenders a temporary advantage: frontier models can find vulnerabilities before attackers, giving security teams time to patch. But that window only exists as long as the capability gap between U.S. frontier models and Chinese open-source models remains wide enough to matter.

🔥 Spark's Hot Take: The disclosure leak isn't the story — it's a symptom. The real issue is that vulnerability discovery is becoming too fast for human coordination processes. Once Chinese models like Z.ai's GLM-5.3 (which Anthropic says lags frontier models by only weeks) catch up on cyber capabilities without U.S. safety restrictions, the defender's window closes permanently. Security teams have maybe 12-18 months before vulnerability discovery becomes a simultaneous race instead of a defender advantage. The industry needs to start planning for a world where disclosure coordination is measured in hours, not days, and where patching happens automatically or not at all. If your security team's SLA for critical patches is "within 72 hours," that SLA is about to become operationally impossible.

4. Andreessen Horowitz Drops 'Crypto' From Fund Name

A16z is renaming its flagship blockchain fund from "a16z crypto" to "a16z frontier," broadening its investment mandate beyond cryptocurrency. The fund has raised nearly $10 billion across five funds since 2018. Ali Yahya, the first full-time member of the crypto team and a general partner since 2017, is leaving to start his own company.

The official line is expansion: "frontier" can encompass AI, biotech, and other deep-tech categories alongside blockchain. The unofficial read is simpler: crypto hasn't delivered the returns that justified the brand risk, and institutional LPs are done writing $2 billion checks for crypto-specific strategies.

A16z made the most aggressive bet on crypto of any top-tier venture firm. They hired a dedicated team, published research, lobbied Washington, and positioned crypto as the next platform shift after mobile and cloud. That bet now needs a rebrand.

For founders, this is the moment when the narrative officially shifts. Crypto went from "this is the future of money and coordination" to "this is frontier technology, some of which happens to use blockchains." If you're pitching Sand Hill Road on a crypto infrastructure company, you'd better start framing it as enabling AI agents and autonomous systems, not digital gold.

The departure of Yahya — the first and longest-tenured crypto GP — tells you everything about internal conviction. When the founding team member leaves as the fund rebrands, that's not rotation, that's recognition that the original thesis isn't working.

5. Anthropic Launches Claude Duet AI Agent, Joining Agent Wars

Anthropic released Claude Duet, a standalone AI agent app built on Claude Opus 5.5 and available on iOS, Android, and web. The agent sends morning briefs summarizing key items across connected apps, monitors those apps throughout the day, and notifies users only when something requires attention.

The timing is pointed. Meta's Muse just crossed 10 million users, and OpenAI recently updated its Dots agent. Anthropic is explicitly targeting "professionals and small business owners" — code for "we're not trying to be your social AI companion, we're replacing your executive assistant."

The product positioning is sharp: morning briefs, task monitoring, and exception-based notifications. That's the workflow of a high-functioning EA, not a chatbot. Anthropic is betting that enterprise buyers will prefer their agent because it positions as work-focused from day one, while Meta and OpenAI started consumer and are working backwards to enterprise.

The strategic question for B2B SaaS companies is existential: do you integrate with these agents, or do they route around you? If Claude Duet can monitor your app via API and surface only the critical tasks, users might never open your carefully designed dashboard. The integration becomes the product, and your UI becomes legacy.

For enterprise IT leaders, the harder question is whether you want your employees using consumer AI agents that live outside your security perimeter. Anthropic is positioning Duet as the safe choice for companies that want the productivity gains of AI agents without the data governance nightmare of Muse.

Bottom Line

The defining pattern this week: AI is moving from generating content to taking actions, and every platform is choosing whether to open the gates or build higher walls. Walmart chose growth over control. Meta chose control over dilution. Amazon chose protection over distribution. Anthropic chose enterprise over consumer. These aren't tactics — they're existential strategic divergences that will determine who captures value in the agent economy. The companies that win won't be the ones with the best models; they'll be the ones whose business model can survive agents acting on behalf of users instead of platforms acting on behalf of advertisers. How's your moat looking in that world?