What would an agent pay for?

The question asked most often under the video was ‘what would I actually sell?’ This lesson answers it with a test you can run on your own idea, real sellers that pass and fail it, and a scorer at the end.

“Still not even one example?” was one of the comments under the video this course grew from. It is a fair complaint, so this lesson starts with examples and builds a test out of them.

The test has two parts.

What would an agent pay for?

1. Would an agent pay for this? It pays when you have something it cannot get free at the same quality: fresher, more local, cleaner, or yours alone.
2. How often? Income is price per call times the number of calls. Something an agent needs every hour is worth far more than something it needs once.

An agent does not decide to pay on a whim. A person or a company sets it up, puts money in its wallet and decides which sources it may use. So the real question is whether that person would point their agent at you instead of at a free source.

Rainfall from a farm gauge

The most-liked idea in the comments (154 likes) came from a viewer who said they are a software engineer and a farmer: log rainfall through the day, with a location and a timestamp on every reading, and let weather agents pay to read it.

Run it through the test.

• It is fresh. Readings change through the day, so an agent has a reason to come back.
• It is structured. A reading with a place, a time and a unit is something a program can use without guessing.

Rainfall from a farm gauge

• It is yours. You own the gauge, so you have the right to sell what it records.
• Free alternatives are the hard part. If a national weather service already publishes rainfall near you, a weather company will use that. Your readings earn money where there is no free station nearby, or where you measure something finer, such as rain every five minutes in a valley the official network misses.

Rainfall from a farm gauge

Now the “how often”. Say five weather agents check your gauge every 15 minutes. That is 480 calls a day. At a fifth of a cent a call ($0.002), you take about $29 a month before costs.

Rainfall from a farm gauge

One gauge earns pocket money. A set of gauges in places nobody else covers, sold to a firm that needs that coverage, is closer to a business.

Messy data made clean

Another viewer put the idea well: look for “messy data that you can structure where it would be more efficient for an agent to pay your site than structure the data itself.”

Here is a made-up case. The planning applications for your area appear as scanned PDFs on a council website, a few each week. An agent working for a builder could open every PDF itself, every day. Or it could pay you a few cents for a clean list with the address, the postcode, what is proposed, the date and a link. You sell the cleaning. The agent saves time and computing cost.

Messy data made clean

Check two things before you build it. Read the licence, because some public data may be reused and resold and some may not. And search for anyone already giving the clean version away.

Niche tools and MCP tools

You do not have to sell data. A tool that does one job well can work too: a check, a calculation, a file conversion. x402 can charge per call for an ordinary web address, or for a tool inside an MCP server, which is how AI assistants connect to outside tools. Coinbase’s catalogue of paid services lists both kinds.

Niche tools and MCP tools

Cred Protocol is a real example. It sells wallet credit scores to agents, at $0.01 for a score and $0.07 for a full report, with good documentation and an MCP server. On paper it passes most of the test. From January to October 2026 its x402 sales on Base came to $0.88 in total, from four buyers. The tool was ready, and the buyers did not come.

Niche tools and MCP tools

One of the few sellers in the research where agents visibly buy something real is Laso Finance. It sells prepaid cards, gift cards and payouts. In the 30 days to 2026-10-11 it took $1,487 on Base, and most of the money came from orders of $5 to $30 across more than 50 wallets. The researchers called it the most natural-looking mix of buyers they found. An agent cannot download a gift card for free.

Ideas that fail, and why

A finished guide or a hobby booklet fails on “how often”. Once an agent has read it, it has what it needs. You can let people read free and charge agents (lesson 4 covers how), but expect one small payment per agent, not a stream.

A course, or anything made for people to learn from, fails because agents buy inputs to a task, like data, a check or a conversion. One commenter picked a social media course as the example of what this is wrong for, and they were right. Sell courses to people in the usual way.

Ideas that fail, and why

A wrapper around a free service fails on free alternatives. Say you resell company lookups from Companies House. Companies House gives that data away free through its own API, so any agent’s owner will go there. A wrapper earns only if it adds something real, such as your own checks or scoring on top.

Ideas that fail, and why

Data copied from someone else’s site fails on rights. No score anywhere else makes up for that.

Popular data that many sellers already have fails on price. Crypto news is the example in the research. Gloria AI charges $0.03 for a news call, and another seller, Otto AI, charges $0.001 for a similar one. In the 30 days to 2026-10-11, Gloria earned $1.29 from x402 on Base.

A listing brings few buyers

On 2026-10-14 Coinbase’s catalogue of paid x402 services, called the Bazaar, held 35,073 entries. The median entry had 2 calls in the previous 30 days, and 89% had five or fewer. The researchers think some of those calls are the catalogue checking its own listings.

So a strong idea still needs a buyer you can name: a weather firm, a builder, a research desk whose agent would call you on a schedule. If you cannot name one, ask around before you write any code.