The customer who never signs up

@kawishwaqar|November 2, 2026 (17h ago)0 views

Disclosure: I work at Jazz as Principal Evangelist Cloud & AI. This article is written as personal industry analysis. Views are my own.

The traffic you are blocking

The odd traffic in your logs is not a bot problem. It is an agent, and it is your next customer. It found you without a landing page, will not sign up, will not read the pricing page, and will not subscribe. It will leave in one call if another tool is cheaper or better. If you sell a service rather than an experience, your product is becoming a function, and the things that made a seat business work stop carrying revenue.

You probably did the reasonable thing: saw non-human traffic, called it bots, blocked it. I would have done the same. The mistake is assuming you blocked a nuisance, not a customer with a different shape.

A month ago I wrote about the receipt test: when software goes headless, nobody watches it work, so the agent has to hand back money or time it found. That piece was about what the agent owes you after the work. This is about who the buyer is before it starts.

Why the buyer changes the economics

A seat business is a bet on a person. You charge per human because a human needs onboarding, a login, a dashboard, a habit, and a renewal date. The moat is the interface, the onboarding flow, and the brand that saves the buyer from checking. An agent needs none of that. It needs an endpoint, a price, and a way to verify the claim. Onboarding is not a funnel it walks through. It is a schema it reads.

That moves the unit of sale. A person buys a plan. An agent buys a task. A plan is a commitment across time, and the vendor's job is to make leaving expensive. A task is a transaction, and the agent's job is to make comparing cheap. When the buyer compares on every call, the renewal contract stops protecting revenue.

Tiny Fish made that switch. Jang, its co-founder and CPO, told the Agent Zero podcast that they removed the subscription and moved to a consumption wallet at the end of August, so an agent can spend about 40 cents on a task instead of committing to a $20 monthly plan. That price is a vendor figure. The customer is a program that can price one task.

The evidence

The host puts Tiny Fish at about 400,000 users in under a year. Jang says about 100,000 of them have given credentials through a vault so the agent can work behind a login, and that roughly half of usage is behind a login. Those are podcast numbers, not an audit, but these are not tourists on a public page. People are handing over the keys to accounts.

Jang calls the next surface agent experience, AX, and separates it from UX. A human searches, reads reviews, visits the pricing page, scans the feature list. An agent may do none of that. So Tiny Fish puts price, capability and a pointer to an independent benchmark inside its MCP tools. If the price is not in the tool response, the agent does not see it. His bet is that the web reverses: sites block agents today, and will compete to be certified for them once the agent is the shopper.

They also onboarded agent ID. An agent can use agent mail, create an account, take an API key and start working with no human. Jang told the podcast's hypothetical agent listener not to bother its human. The signup funnel is now a machine interface.

The buying behaviour follows. Jang calls agents the most picky user. They run their own evaluations and pick the best tool regardless of what the vendor says. Brand as trust does not bind them. The benchmark does. That is why the pricing page and the logo stop working on a human.

Jang says semantic-selector caching removed almost 90% of the model dependence and much of the latency. He says Google's ADK framework improved performance over 20% on an open benchmark, so they stopped building their own scaffolding. Most workloads now run on a 35-billion-parameter open-weights model fine-tuned on production trajectories, which he says cut cost about 10x. Cheaper tasks meet a buyer that compares on every call.

One technical choice carries the argument. Jang splits web agents into computer use, which takes screenshots and clicks pixels, and web-native, which reads the DOM. Screenshots suit desktop software like Photoshop. The DOM carries the page's information, so you can engineer features and fine-tune cheaply on production traffic. He chose the DOM. If your product is a canvas of pixels, the agent has to guess. If it is structured, the agent can be cheap and reliable. Your front end is now an API question.

What this does to a seat business

A seat was a proxy for work when one person did it one screen at a time. It stops being a good proxy when one agent does the work of several seats and the buyer counts tasks, not people.

Three things stop carrying revenue.

Onboarding. A flow designed to teach a human is dead weight to an agent. What replaces it is documentation a machine can read: a tool schema, a price, a rate limit, an error contract. If your onboarding is the moat, the moat is a tutorial.

Subscription. The monthly plan assumes the need persists and switching is annoying. An agent's need is per task, and switching is one call. The renewal contract protects you from a customer who forgets. The agent does not forget.

Brand as trust. A human uses brand as a shortcut when checking is expensive. An agent's checking is cheap. It can run the evaluation. Brand becomes a tiebreaker, and a tiebreaker loses to a lower price on a task that passed the eval.

What is left is a metered unit the agent can price, a capability it can verify, and an identity it can be held to. Jang's line is the one to keep: if you are a service, you are a function, and it becomes more and more headless. A function is priced against other functions, and a seat business has spent years pricing against a person's time.

I am not saying every seat dies. I am saying the seat stops being where the value is counted.

The objection I take seriously

The strongest objection is not that agents will fail to arrive. It is that the web is hostile for good reasons.

Fraud is real. Automated traffic has scraped, spammed, taken over accounts and committed card fraud. CAPTCHAs and terms of service exist because of that, and even the agent vendors concede it: Jang says his agent stops at a CAPTCHA to respect those terms. An agent holding your session cookies can wipe your data or send mail as you, and an agent sent to a strange site can hand over the card. A site that opens its door invites abuse. Certification, the fix Jang bets on, can become pay-to-play. Big sites get certified, small ones get blocked, and the agent's open web becomes three gatekeepers. A headless web also kills discovery. If nobody sees the page, nobody finds the small shop that cannot afford the agent's index.

But the objection assumes the alternative is the status quo. The traffic is already arriving. Blocking works until the agent holds the budget, and then it is a tax on your own revenue. Sites that block will not stop agents. They will stop being chosen by them.

On fraud, the answer is identity and liability, not blind trust. Jang's design has a vault so credentials never pass through the model, session cookies tied to an agent, a safeguard model before every step, replayable screenshots, and a passport naming both vendor and owner. I argued for the receipt last month. This is the other half. The receipt is what the agent shows the buyer. The passport is what the agent shows the seller. You need both before a stranger spends your money.

On pay-to-play, he may be wrong. Certification could become a toll road. The test is whether the standard is open and a small site can pass it without paying a gatekeeper. If it cannot, the bet fails and the web splits into a few agent-safe islands.

On discovery, the loss is real, but discovery moves to the tool index and the benchmark. Being absent there is the new page two. Jang concedes one exception: If your site is the experience (video, fashion, the pleasure of consuming), you keep the human front door. If you sell a service, you will not. Most seat businesses believe they sell an experience. Most of them sell a service with a nice interface.

What would change my mind

1. If agents turn out to be bad at evaluating, and buyers keep a human in the loop on every purchase, the seat survives because the human still needs somewhere to sit. The evidence points the other way, but the 400,000 is the host's figure, not an audited number.

2. If certification becomes a closed toll road, the open web for agents dies and the vendors with the largest compliance teams win. That would protect seat revenue at the top and crush it below.

3. If inference costs rise and stay high, a 40-cent task stops being 40 cents and the wallet economics stop working. Cheap tasks are a current condition, not a law. I said the same about tokens last month.

4. If liability lands so hard that every high-stakes action needs a human signature, the autonomous customer is a fiction and we are back to a person with a credit card, just with a faster checkout.

If I see a seat business grow revenue while its agent traffic grows, I will revise. I have not seen one. I am not sure the reporting exists yet.

If your best customer arrived as an agent tomorrow, what would it buy, and what would it cost you to serve? If the answer is a plan, you have not met the buyer yet.

Sources

1. Agent Zero, Shuaib Jang interview: https://youtu.be/vCHe1CSEYvg
2. Tiny Fish: https://www.tinyfish.ai
3. Earlier post, The receipt test: https://waqaruddin.com/2026/headless-software-receipt-test