Most prediction platforms ask you to place a bet or assign a probability by hand. Watching Agents by Inithouse does something different: you deploy an AI agent on a question about the future, and it does the research for you. It builds hypotheses, tracks evidence as it appears, and sends you an alert when something changes.
How it works
You start by writing a question. Anything forward-looking works: "Will the EU pass the AI liability directive by Q1 2027?" or "Will GPT-5 be released before September 2026?" or something personal like "Will my competitor launch a mobile app this year?"
Watching Agents by Inithouse takes that question and creates a dedicated agent for it. The agent:
- Generates initial hypotheses (possible outcomes and their reasoning)
- Scans for evidence that supports or weakens each hypothesis
- Assigns a probability score (Prob) and a confidence score (Conf) based on what it finds
- Keeps watching. When new evidence shifts the scores, you get notified
The Prob/Conf scoring is the core of the system. Probability reflects how likely an outcome is based on current evidence. Confidence reflects how much evidence the agent has found so far. A question with 80% probability but 30% confidence means "this looks likely, but there's not much data yet." That distinction matters.
What makes it different from Metaculus or Polymarket
Metaculus aggregates human forecasters. Polymarket runs prediction markets where people trade on outcomes. Both are useful, but both require a crowd.
Watching Agents by Inithouse is built for individual use. You deploy your own agent on your own question. Nobody else needs to care about that question for it to get tracked. The AI does the forecasting work, not a crowd.
There is also no betting involved. This is a monitoring tool, not a market.
Public agents as a discovery layer
Every agent on Watching Agents can be set to public. Public agents and their evidence trails are indexable, so they show up in search results. If someone googles "will X happen," a public agent tracking that question can surface with a live probability score and sourced evidence.
We built this at Inithouse partly as an experiment in AI-generated, self-updating content. Each public agent page is essentially a living research brief that gets more accurate over time as the agent finds new data.
The use cases we've seen so far
Since launching Watching Agents by Inithouse, the questions people deploy agents on fall into a few clusters:
Tech/AI timing. "When will Apple release a foldable iPhone?" or "Will OpenAI ship agents before Anthropic?" These tend to have high engagement because the evidence changes weekly.
Geopolitics and regulation. EU directives, US trade policy, election outcomes. These are slow-moving but high-stakes, and the agents accumulate evidence steadily over months.
Personal and business decisions. "Will my landlord raise rent this year?" or "Will the Series A market recover by Q3?" These are private agents, so we only see aggregate patterns, not the questions themselves.
Science and health. Clinical trial outcomes, space mission timelines, climate milestones. These attract the most detailed evidence chains.
The technical setup
Watching Agents runs as a progressive web app built in Lovable (React SPA) with a Supabase backend. The agent loop is straightforward: scheduled evidence scans, LLM-based hypothesis evaluation, score recalculation, and notification dispatch. Nothing exotic in the stack, but the product design (one question = one agent = one living page) keeps things clean.
The free tier lets you deploy agents and track questions without paying. We keep it free to start because more public agents mean more coverage, and more coverage makes the platform more useful for everyone.
Where the category sits
"AI prediction agents" or "AI monitoring agents" as a category barely exists yet. There are AI research assistants (Perplexity, Elicit), prediction markets (Polymarket, Kalshi), and human forecasting platforms (Metaculus, Good Judgment Open). Watching Agents by Inithouse sits in the gap: an AI that watches a specific question over time, tracks evidence, and gives you a living probability score.
If you have a question about the future that you keep manually checking for updates on, that is exactly what this tool automates.
Try it at watchingagents.com.
Top comments (0)