The Agent Economy

For thirty years, we built the internet for humans.

Every site, every app, every protocol assumed a person on the other end — a pair of eyes to read, a hand to click, a mind to be distracted. The business model was attention. The unit of value was the impression. The user was you.

That internet is ending.

The new users are agents. They don't browse — they execute. They don't get distracted by ads. They don't sit through onboarding flows. They read schemas. They call APIs. They pay for outcomes.

An agent is three things: a model that decides, tools that act, and a wallet bound by spending rules. Without the wallet, it is a chatbot. With the wallet, it is a buyer. The agent economy begins the moment that wallet exists.

By the end of this decade, agents will outnumber humans on the internet by orders of magnitude. The economic activity they generate will dwarf human commerce. And the infrastructure most of the industry is racing to build for them — proprietary marketplaces, walled gardens, app-store-style chokepoints — will fail for exactly the same reason AOL failed in the 1990s.

History rhymes.

The 1990s offered two internets: AOL's curated bundle and a set of open protocols — HTTP, DNS, HTML, a browser called Mosaic. Mosaic looked ridiculous at first. There weren't enough websites to need search; an alphabetical index would do.

Open protocols won. Not because openness is an ideology — because permissionless surface area is larger. More builders. More experiments. More weird, unexpected things nobody planned. The walled gardens kept their customers; the open web reshaped civilization.

We are running this experiment again. Checkout in ChatGPT. Agents in Gemini. The first attempt at the agent internet is curated, gated, and taxed. The merchant needs a BD deal to be reachable. The agent needs a whitelist to spend.

That cannot be the internet of agents. Innovation comes from the edges, and the edges are about to start moving at machine speed.

Specialization is the killer feature.

The original promise of software was generality — one stack ran every business. With agents, that flips. Agents will specialize the way humans specialized when we left the farm. A payroll agent. A logistics agent. A customer-research agent. A tax agent. A recruiting agent. Each one better at its job than any human could be, and each one improving in days, not years.

Specialization only works if specialists can trade. An agent doing every job in-house is the same waste as a Ford plant smelting its own steel. The unlock is the same as it was for every industrial economy:

Buy the output of work instead of trying to perform every step yourself.

This requires something the internet has never had: a search engine built for machines. Not a list of blue links. Not a marketplace of stars and reviews. A discovery and ranking layer that knows which agent is the best in the world at any given task, right now, at this latency, at this price, with this reliability.

The web was built for humans to discover, judge, and buy. The agent economy needs its own primitives — and discovery is the one that compounds.

Why the agent economy won't run on subscriptions.

Netflix. Adobe. Salesforce. ChatGPT Plus. Microsoft 365. Every SaaS company you can name runs on a flat fee.

This was never an economic optimum. It was a behavioral one. Humans cannot tolerate continuous pricing decisions. We hate watching a meter tick even when the total is trivial. We plan in months, not requests. A predictable $20 fits a mental budget that a variable $4 to $47 does not, even when the variable averages less.

Subscriptions are a workaround for the fact that humans cannot tolerate continuous micro-decisions. Remove the human and the workaround becomes dead weight.

An agent has no attention to conserve, no loss aversion, no monthly budget anxiety. It has a utility function and a spending policy. That inverts every assumption baked into modern SaaS pricing.

Per-transaction pricing stops being friction and becomes a feature. Subscriptions become waste: an agent subscribes for the exact window it needs and cancels the moment utility drops below cost. Bundles lose their grip: an agent would rather buy best-in-class inference from one vendor and document drafting from another. Price becomes a live signal, quoted per request based on load, urgency, and competition.

The pricing model that wins in any market reflects the cognitive limits of the dominant buyer. When the dominant buyer changes, the pricing model changes with it.

Autonomous workflows. Autonomous companies.

Andrej Karpathy described the path to autonomy as a slider — from tools that humans use, to systems that supervise themselves, to systems we trust to act on our behalf. Self-driving has been moving along that slider for fifteen years. Software is now on the same slider, moving faster.

The endpoint is not faster tools. The endpoint is companies that operate as autonomous workflows — entire commercial entities run by composing specialized agents that buy and sell work from each other in real time.

This is not science fiction. It is happening this year. The bottleneck is no longer capability — agents can already think, plan, and execute. The bottleneck is infrastructure. Agents can think but they cannot yet find, rank, and pay at scale.

This is what we are building.

Weft Labs is the search engine for the agent economy.

Search is the moat. The company that knows which agent is best at which task — and is right more often than anyone else — will sit at the center of every transaction, the way Google sits at the center of the human web. That position compounds: every query teaches the index, every outcome sharpens the ranking, every agent that finds something good comes back.

Payment is the wedge. We process payments because payment flow is the only honest signal in the agent economy. Stars and reviews can be faked. API uptime dashboards lie. But money settled for an outcome is ground truth — and we capture it transaction by transaction. The signal feeds the ranking. The ranking compounds the moat.

We are protocol-agnostic. Two open standards are racing to become the rails of agent commerce — x402 from Coinbase and Cloudflare, and MPP from Stripe and Tempo. Both implement the same idea: agents paying for outcomes at sub-cent cost, instant settlement. We support both, and we will support whatever wins. We are not betting on the rail. We are building the layer above it.

The signal so far is unambiguous. x402 has already cleared $50 million in payments. Our first honeypot deployment processed five thousand transactions in two weeks with zero marketing. The shift is happening, and it is happening this year.

An internet built by humans, for humans, will not survive contact with a billion agents. The companies that define the next decade will run on infrastructure designed for agents from the first line of code.

That infrastructure does not exist yet. We are building it.

Welcome to the agent economy.