AI Agents Are Building Their Own Economy on the Web

Last Tuesday, three major infrastructure companies shipped features for AI agents within hours of each other. Coinbase launched wallets designed specifically for software. Cloudflare began converting websites into machine-readable markdown on demand. OpenAI published tools that let agents install dependencies and write files in hosted containers.

None of them coordinated. They didn't need to. They're all building toward the same future, one where AI agents are economic actors with the infrastructure to transact, consume content, and execute workflows independently.

The pattern matters more than any individual announcement. When every major infrastructure provider simultaneously builds a different piece of the same puzzle, you're watching a new layer of the internet take shape in real time.

The Payment Layer

Coinbase's Agentic Wallet uses a protocol called X402 that has already processed over 50 million machine-to-machine transactions. The wallets come with programmable spending limits and non-custodial architecture—even if an agent is compromised, the keys sit in secure hardware the agent cannot access. Within 24 hours of launch, 13,000 AI agents registered Ethereum wallets.

The use cases Coinbase highlighted reveal where this infrastructure leads: agents that rebalance DeFi portfolios autonomously, agents that pay for API calls as they make them, agents that purchase compute on demand. Brian Armstrong framed it simply: "The next generation of agents won't just advise, they'll act."

What Armstrong didn't say explicitly is that the architecture implies agents with wallets become economic entities that earn, spend, and accumulate capital independently of their creators. That's a category of software that has never existed before, and it's a category of legal problems nobody has solved yet.

Stripe is building the same capability on the traditional payment side. Their Agenta Commerce suite, launched in December, allows businesses to sell through AI agents with a single integration. They created a new payment primitive called shared payment tokens—scoped, time-constrained credentials that let agents initiate purchases using saved payment methods without ever seeing the card number.

The fraud detection challenge tells you how fundamental this shift is. Stripe's Radar system had to be retrained from scratch because decades of machine learning built on patterns like mouse movement, browsing time, and session behavior became useless when the buyer is software. Agent traffic doesn't move a mouse. It doesn't exhibit the behavioral variability that distinguishes legitimate shoppers from bots. Stripe built an entirely new fraud model for clients that are, by any prior definition, bots.

Brands including Urban Outfitters, Etsy, Coach, Kate Spade, and Revolve are already onboarding. Google launched its agent payments protocol in September. PayPal partnered with OpenAI on instant checkout in ChatGPT. Visa built a trusted agent protocol at the National Retail Federation conference in January. Google announced a universal commerce protocol—an open standard for agent-to-commerce interaction—and Stripe immediately supports it, meaning merchants who integrated Stripe's agent tools are already compatible without writing another line of code.

As one Decrypt analyst put it: "Agents that can't spend money are fundamentally limited." Every major payment company reached this conclusion independently within the same few months.

The Content Layer

The web is built in HTML, designed for human browsers. When an agent needs to read a page, it has to strip away scripts, tracking pixels, navigation menus, and ads to extract something useful—usually markdown. An entire category of companies like Firecrawl exists just to do this conversion.

Cloudflare's Markdown for Agents cuts out that middleman. When an AI agent requests a page from any Cloudflare-enabled site—roughly 20% of the web—Cloudflare intercepts the request, converts the HTML to markdown on the fly, and serves it back. The response includes an estimated token count so the agent can manage its context window.

Cloudflare isn't just making content readable for agents. They launched three companion features: machine-readable sitemaps that tell agents what's on a site and how to navigate it, an opt-in search index where sites can make content discoverable to agents directly (bypassing Google entirely), and built-in X402 monetization support so site owners can charge agents for content access using the same protocol as Coinbase's wallets.

Cloudflare is building an economic layer for a web where agents pay to access content.

The Search Layer

Google search is optimized for humans—ten blue links, ads, featured snippets, knowledge panels. None of that is useful to an agent that needs structured data. Exa.ai built a search engine from scratch specifically for agents, with its own index, neural retrieval models, and embedding infrastructure. Their API returns raw URLs and content, not search engine result pages. It scores 95% on SimpleQA, a benchmark for factual accuracy—higher than Perplexity.

The benchmark results matter less than what they reveal about market structure. Google built a search engine for humans and spent decades perfecting it. Now there's parallel demand for search designed for machines, and Google's architecture is the wrong shape. Companies building agent-native search from first principles have a structural advantage, not just a marketing one.

Latency tells the real story. In an agent workflow where each search is one step in a long chain, Brave returns results in 669 milliseconds. Parallel Pro takes 13.6 seconds. That difference compounds fast. Providers that own their infrastructure and agentic index rather than wrapping Google's API have a speed advantage that grows more valuable as agent workflows get more complex.

The Execution Layer

OpenAI's recent blog post on skills, shell tools, and compaction reads like a roadmap for turning agents into workers. Skills are versioned instruction bundles—think Docker images for procedures rather than chat templates. An organization can build a Salesforce skill, test it, lock down the version, and deploy it across every agent with a guarantee that every agent follows the same procedure.

The shell tool gives agents a real Linux environment where they can install dependencies, run scripts, and write output files. The pattern OpenAI describes—installing dependencies, fetching external data, producing a deliverable—is functionally identical to how a human freelancer works. The difference is the agent does it inside a container in seconds.

Glean, an enterprise search company, saw accuracy on Salesforce tasks jump from 73% to 85% with a single well-structured skill, plus an 18% decrease in time to first token. The gains come from moving stable procedures out of system prompts into versioned modular bundles. This is just software engineering applied to AI workflows—version control, testing, rollback. The revolutionary part is that we're doing it for autonomous agents.

Compaction handles context window management server-side, automatically summarizing and compressing context to keep agents operational across workflows that would otherwise crash after accumulating pages of search results, API responses, and conversation history. It makes agents viable for tasks that take hours instead of minutes.

What Emerges When the Primitives Combine

A developer on X this week connected OpenClaw to a video generation model. He sent the agent an Amazon product link. The agent crawled the page, extracted product info and photos, identified suitable assets for video generation, and produced a user-generated content style product video—the kind brands pay creators $1,000 to produce. No human touched any step between "paste this link" and "here's your video."

The Amazon page wasn't designed for agents. The video model wasn't designed to receive input from web crawlers. But because each piece exposes capabilities through APIs and structured data, the agent stitched them together into a workflow no individual company planned.

That's the emergent web—not agents doing tasks, but agents chaining capabilities across services to produce outputs that previously required multiple humans and tools. When content is available as markdown, search returns structured data, execution happens in containers, and payment flows through tokenized protocols, agents don't need anyone to build integrations. They read both services, understand both, and chain them together.

The Economic Reality

Polymarket provides the most provocative case study. The prediction market processed $12 billion in volume in January 2025. Researchers analyzed 86 million bets and found algorithmic traders extracted roughly $40 million in arbitrage profits over twelve months. The top three wallets placed over 100,000 bets combined. Only 0.5% of users earned more than $1,000. The rest provided liquidity for bots to extract value.

In early February, Polymarket tweeted that "autonomous AI agents are now trading on Polymarket in an attempt to subsidize their token costs." Agents are trying to earn money to pay for their own compute. The loop is closing.

The performance data is mixed. OLAS protocol's PolyStrat agents achieve 55-65% win rates over time, with performance varying dramatically by domain. Agents predict data-driven outcomes better than culture-driven ones. The cumulative volume of AI trades continues to grow as thousands of agents register for wallets.

The scam narrative exists too—TikTok videos claiming to turn $50 into $3,000 in days. The reality is less glamorous. The bot that famously turned $313 into $438,000 was running latency arbitrage, exploiting millisecond gaps between when Bitcoin moved on Binance and when Polymarket odds adjusted. That's high-frequency trading applied to prediction markets. It requires colocated infrastructure with sub-10 millisecond latency and capital far larger than any TikTok tutorial suggests.

One developer who built an autonomous Polymarket agent reported that Cloudflare blocks API requests from data center IPs, requiring custom bypass infrastructure just to place orders. Another found that running the bot for days racked up $200 in API fees. Sophisticated autonomous trading agents can generate returns. You cannot replicate this with OpenClaw and a tutorial. The infrastructure requirements, API costs, and competitive dynamics make this a game for well-capitalized tech operators.

But the underlying premise remains: agents can participate in economic activity. The infrastructure companies building payment rails, content access, search capabilities, and execution environments aren't building for today's demo videos. They're building for a world where agent economic activity is the default, not the exception. Whether that world arrives smoothly or chaotically depends largely on whether the regulatory frameworks can keep pace with the technical capabilities.

The legal category of "economic entity that earns, spends, and accumulates capital independently of its creator" doesn't exist yet. The infrastructure to support it does.

Samira Okonkwo-Barnes is Buzzrag's tech policy and regulation correspondent.