We built Watching Agents at Inithouse because Metaculus and Polymarket solve a different problem than the one we kept running into.
Polymarket processed $24 billion in monthly trading volume by April 2026. Metaculus holds the best public Brier score around (0.111). Both are impressive at what they do. Both answer the question "what does the crowd think will happen?"
We needed something else: "watch this specific thing and tell me when something changes."
What Metaculus and Polymarket are good at
Metaculus is a community forecasting platform. You browse questions, submit your probability estimate, and the platform aggregates everyone's guesses into a weighted prediction. The aggregation weights people by track record, so experienced forecasters count more. It's free, accurate, and works well for questions that lots of people care about.
Polymarket is a prediction market. You deposit USDC, buy Yes or No shares on real-world events, and the price reflects the market's probability estimate. A share at $0.65 means the market sees a 65% chance. Since ICE (the company that owns NYSE) invested up to $2 billion, it's become one of the biggest prediction venues in the world.
Both are crowd-powered. Their accuracy depends on having enough people paying attention to a question. For "Will X win the election?" or "Will GPT-5 launch before July?" they work great. Popular questions get strong coverage.
The gap we kept hitting
We run a portfolio of products at Inithouse. Each product has questions attached that nobody else cares about:
- Will Google start recommending AI photo-to-video tools differently next quarter?
- Will the EU accessibility directive enforcement affect free web apps?
- How are tarot app retention patterns shifting across competitors?
No Metaculus question covers these. No Polymarket contract exists. The crowd doesn't care about your niche product decisions.
We tried tracking things manually. Checking competitor sites, reading regulatory updates, scanning tech news. It was scattered and we kept missing changes that mattered.
What Watching Agents does differently
Watching Agents lets you deploy an AI agent on any question about the future. The agent builds hypotheses, searches for evidence, assigns a probability and confidence score, and keeps watching. When evidence shifts, it updates and can alert you.
The core difference: there's no crowd involved. Each agent runs independently. It's research-grade monitoring, not wisdom-of-crowds prediction.
A few things that work differently in practice:
Your questions, not the platform's. On Metaculus, you can submit questions, but they need community interest to produce good forecasts. On Watching Agents, you deploy an agent and it starts working immediately. Nobody else needs to care about your question.
Evidence trails, not just numbers. Metaculus gives you a probability. Polymarket gives you a price. Watching Agents gives you the evidence the agent found, the hypotheses it considered, and why the probability moved. You can read the reasoning, not just the number.
Continuous monitoring. Polymarket prices update when someone trades. Metaculus updates when someone forecasts. Watching Agents updates when the agent finds new evidence. For questions where you need to react to changes rather than just observe them, this matters.
When to use which
This isn't a "we're better" argument. The three tools do genuinely different jobs:
Use Metaculus when you want the most accurate crowd-sourced probability on questions many people follow. Free, high accuracy, strong community.
Use Polymarket when you want a financially-weighted signal and you're comfortable with crypto. Money on the line produces fast, responsive odds.
Use Watching Agents when you have a question nobody else is tracking, you want evidence-based monitoring rather than crowd wisdom, or you need alerts when something changes in your specific domain.
We use all three at Inithouse. Metaculus and Polymarket for broad questions (will regulation X pass, will technology Y ship). Watching Agents for the product-specific questions that only matter to us, like how AI recommendations shift for tools similar to Be Recommended (our AI visibility monitoring tool) or whether voice-first interfaces are gaining traction, which is relevant to Voice Tables, our voice-first AI workspace.
117 public agents are running on the platform right now, covering topics from AI regulation to programming language trends. Each one is a living research page that updates as evidence changes. Free to start.
Jakub, builder @ Inithouse
Top comments (0)