By James Aspinwall
Source: The $285B Sell-Off Was Just the Beginning — The Infrastructure Story Is Bigger
The web is forking. Not dramatically, not with a press release — but quietly, structurally, and probably irreversibly. In a single week, Coinbase launched Agentic Wallets, Cloudflare shipped Markdown for agents, and OpenAI published tools for skills, shell access, and context compaction. Stripe, Google, Visa, and PayPal are all making similar moves. The common thread: AI agents are no longer tools. They are clients — economic actors that need their own infrastructure.
Nate’s video lays out the thesis clearly. One branch of the web stays optimized for humans: HTML pages, visual layouts, ads. A parallel branch is emerging for software agents that need structured data, payment primitives, and execution environments. This is the agentic web, and the companies building it have enough scale to make their design choices into de facto standards.
Agents as Economic Entities
Coinbase’s Agentic Wallets give agents non-custodial crypto wallets with programmable spending limits, session caps, and gasless trading on Base. Keys sit in secure hardware the agent cannot directly access. Within 24 hours of launch, 13,000 AI agents had registered wallets on Ethereum. Software has never before operated as an economically autonomous actor at scale, and the legal and governance implications are entirely uncharted.
On the fiat side, Stripe’s Agent Commerce Suite lets businesses sell through AI agents using shared payment tokens — scoped, time-bound credentials that allow an agent to charge a saved card without ever seeing the number. Stripe had to retrain its entire fraud detection system because the legacy signals (mouse movement, browsing time, device fingerprints) are meaningless when the buyer is software.
Content, Search, and Execution Go Agent-Native
Cloudflare’s Markdown for agents converts any Cloudflare-served page into clean markdown when requested by an AI, stripping scripts and clutter and providing a token count header so agents can manage context windows. They extend this with LLM.txt/LLMs-full.txt sitemaps, an AI index accessible via MCP and APIs, and built-in X42 monetization so sites can charge agents directly for content. The message is explicit: agents are first-class clients of the web, not bots to be blocked.
Search is splitting too. Human-optimized engines focus on SERPs, snippets, and ads. Agent-native engines like Exa.ai return raw URLs and content with APIs that chain searches, parallelize queries, and optimize for low latency. Benchmarking shows agent-first engines that control their own index deliver structurally lower latency and higher accuracy than wrappers around legacy search — and that advantage compounds across long agent workflows.
OpenAI’s contributions define how agents actually work in this new web:
- Skills are versioned, mountable instruction bundles — more like Docker images than prompts. Organizations define standard operating procedures (“Salesforce skill”) and deploy them across agents with version control and rollback.
- Shell tool gives agents a real Linux environment to install dependencies, run scripts, and write files — mirroring how a human freelancer executes a project, but inside a container.
- Compaction manages context bloat by summarizing and compressing history so long-running workflows don’t crash on token limits.
Emergent Workflows
When you combine wallets, search, content access, payment rails, and execution, agents become full-fledged economic actors rather than glorified autocomplete. The video highlights an example where a developer wired OpenClaw to a video model (Kance 2.0 in Chatcut): the agent ingests an Amazon product URL, crawls the page, extracts assets, selects suitable media, calls the video model, and outputs a UGC-style product video. No human in the loop between “paste link” and “here’s your video.”
For the creator economy, this means repeatable content patterns — product videos, descriptions, social posts, email campaigns, comparison writeups — generated at near-zero marginal cost in minutes. The infrastructure companies are clearly building for a future volume that doesn’t yet exist.
Agent Trading and the Limits of Retail Bots
Polymarket serves as a case study. The platform saw $12 billion in volume in January 2026, with algorithmic traders extracting roughly $40 million in arbitrage profits over 12 months. Polymarket itself has acknowledged that autonomous AI agents are now trading there to earn funds to pay for their own compute — closing the economic loop.
But Nate draws a sharp line between serious latency-sensitive arbitrage and TikTok-style “turn $50 into $3,000” fantasies. The famous bot that flipped $313 into $438,000 was doing high-frequency latency arbitrage with colocated infrastructure and sub-10ms latency. Real-world attempts to run autonomous Polymarket agents ran into Cloudflare blocks on data-center IPs and hundreds of dollars in API fees in just days. Meaningful trading agents are tools for well-capitalized operators, not casual retail users.
Security: Agents as Adversaries
Every primitive that makes agents more capable also expands the attack surface. Wallets can be drained by malicious skills. Shell access can run arbitrary code. Search can be steered to adversarial prompts. Cloudflare-served markdown can deliver poisoned content at machine speed.
The security responses are converging on the same assumption: the agent must be treated as a potential adversary, not a trusted employee. Ironclaw sandboxes each tool in WASM. OpenAI enforces strict network allow-lists and container isolation. Coinbase uses enclave-based keys and programmable spending limits.
The deeper tension: infrastructure assumes fully autonomous agents with zero human oversight, while human comfort sits closer to a 70/30 rule where people want to retain substantial control. This trust gap between what infra enables and what people are willing to delegate will be a central tension in AI for the next several years.
The Mobile-Web Analogy
Nate compares the current moment to 2007, when the iPhone forced a decade-long rebuild of the web for mobile. Same underlying infrastructure, but a different interface layer — responsive design, app stores, GPS services, tap-to-pay — that enabled new giants like Uber and Instagram.
Today’s fork is similar, except the new client has no screen at all. It needs structured data, machine-readable content, tokenized payments, programmatic search, and execution environments instead of pixels and scroll animations.
If the mobile fork created trillion-dollar firms tied to mobile primitives, the agent fork will spawn businesses that simply could not exist on the human web. The shape of those businesses depends on how well we embed trust into the new primitives — payments, security, and governance — for agents that are becoming economic actors in their own right.