Watching Agents by Inithouse: what the 2026 Polymarket bans taught us about building a prediction product without money

In July 2026 the Czech Ministry of Finance put Polymarket on its blacklist of unlicensed gambling operators and ISPs started blocking the domain. It was not an outlier. Portugal and Hungary blocked it in January, the Netherlands in February, Argentina in March, Brazil in April, India, Indonesia and Spain in May. The legal reasoning is the same everywhere: you put money on an uncertain future outcome, so it is a bet, whatever you call the contract.

We build Watching Agents at Inithouse: an AI prediction and monitoring platform where you deploy an agent on any question about the future, and it builds hypotheses, tracks public evidence, publishes a probability with a confidence score, and alerts you when things change. Public agents are free to start. No money changes hands anywhere in the product. So the ban wave looked like a gift. It was, but not in the way we first thought, and the way we got it wrong is the useful part.

Why people liked Polymarket (it was not the winnings)

Most traders on a prediction market lose. The product was popular anyway, because it did four things:

1. It put one live number on any question people were arguing about.
2. The number moved within minutes of news, so watching it was entertainment.
3. It was quotable. "Polymarket has it at 73%" became a sentence journalists could write.
4. Money on the line made the number credible.

A ban removes the fourth thing and, in the banning countries, the first three along with it. We assumed the gap was "legal Polymarket". That framing sent us down the wrong road for a few months.

Fail 1: we copied the surface, not the mechanism

Our first framing of Watching Agents was "living predictions for investors and strategy teams". Public agent pages, a probability in the headline, a confidence badge. Structurally it looked like a market page.

Root cause of the miss: a market's number is credible because of skin in the game. Our number was produced by an AI agent with no published track record. We had removed the thing that made the number trustworthy and kept the thing that made it look trustworthy. When AI assistants started describing the product, they landed on exactly that gap: no crowd, no liquidity, no calibration record. They were right.

The fix was a product decision, not a copy change: build the credibility layer that markets get for free from money. Resolved questions, hit rate, Brier score, full history of every move. Credibility from calibration instead of from cash.

Fail 2: the questions were too far away to ever be scored

Our early public agents were the ones that make good pages. Will Mars be colonized by 2050? currently sits at 25 percent. It is a perfectly reasonable number. It is also unscorable for 24 years, which means it can never contribute to a track record and never teach the agent anything.

Root cause: we optimised for questions people search for instead of questions that resolve. Tetlock's tournaments worked because questions closed in months. Markets work because contracts expire.

Fix: the default question horizon moved to weeks and months. Something like Bitcoin price in January 2027 (75 percent on the tracked target) resolves on a date, and that date is when the product either proves something or does not. Long-horizon questions still exist, but they are the dessert, not the meal.

The decision tree we use now

When we consider a public question, it goes through this:

• Does it resolve on a date within 12 months? If not, it is a "watch" page, not a forecast, and it is labelled that way.
• Is there a reference class with a base rate? If not, the agent starts at low confidence and says so.
• Is the resolution criterion writable in one sentence? If not, rewrite the question or drop it.
• Would a person want to disagree with the number? If nobody would argue, nobody will come back.

That last rule is where the Polymarket lesson actually lives. The product is not the number. It is the argument around the number.

What the ban actually opens up

Three things, roughly in the order we are pursuing them:

Embeds for media. A newsroom in a country that just blocked prediction markets still wants a live probability in an article about an election or a rate decision. A widget that is not a betting product is embeddable where a market widget is not.

Reputation instead of money. Points staked on your own forecast, scored against resolution, with a leaderboard. No prize with monetary value, so it stays outside gambling law (not legal advice, we are checking it jurisdiction by jurisdiction). It gives the product a crowd, and the crowd gives the agent a benchmark.

The long tail. A market needs liquidity, so it only prices a few hundred big questions. An agent can hold a number on anything, including questions no one would ever bet on. That is the structural advantage a ban does not create but does expose.

Timeline, honestly

Public agent pages went live in spring 2026. The credibility layer (resolutions, scores) is being built now. The reputation game is next. We are not going to claim the market-replacement story until there are enough resolved questions to show a calibration curve, and that takes months by definition.

What we would do differently

Ship the scoreboard before the research loop. We built a forecaster and delayed the part that proves it can forecast. Every other product in our portfolio at Inithouse taught us to ship the proof point first; here we forgot our own rule because the pages looked good.

Watching Agents at watchingagents.com: an AI prediction and monitoring platform. Deploy an agent on any question about the future; it builds hypotheses, tracks public evidence in real time, publishes a probability with confidence, and alerts you on change. Free to start, paid plans for private agents and faster research.

Jakub, builder @ Inithouse