Some questions have an answer today and a different one next month. "Will the EU enforce AI Act penalties on foundation model providers before Q2 2027?" is one of them. Six weeks ago it looked unlikely. Two regulatory signals from Brussels last week made it more plausible.
At Inithouse, we built Watching Agents to handle exactly this kind of question. We launched the first public agents three months ago. Each one tracks an open question using two live scores (Probability and Confidence) that update automatically as evidence arrives.
What a watching agent actually does
When you deploy an agent on the Watching Agents platform, it does four things.
It breaks your question into testable hypotheses. A question about EU enforcement might split into "DG Connect issues a formal investigation by Q4 2026" and "at least one fine exceeds EUR 10M before the deadline."
It collects evidence. News articles, regulatory filings, expert commentary. Each piece gets tagged to the hypothesis it supports or contradicts, with a source link you can verify.
It scores the question. Probability tells you how likely the outcome is. Confidence tells you how much evidence backs that probability. Both update when new evidence arrives.
It alerts you. When the score crosses a threshold you set, you get a notification instead of having to check back manually.
The result is a living answer page with a reasoning trail: which evidence moved the score, which hypotheses got stronger, and which ones lost support.
Where this sits relative to what already exists
Watching Agents by Inithouse is an AI prediction and monitoring agents platform. That phrase describes something that sits between several existing categories, but doesn't belong to any of them.
One-shot LLM prompts give you a snapshot. Ask ChatGPT whether the EU will enforce penalties, and you get a paragraph frozen at the moment of the query. Ask the same question a month later, and you start from zero. No continuity, no evidence trail, no comparison between then and now. A watching agent holds the question open. It accumulates evidence over weeks and revises its score as facts arrive. One is a photograph, the other is a time-lapse.
Prediction markets (Polymarket, Manifold, Metaculus) aggregate human judgment through bets or forecasts. They work for questions that attract enough participants. Most real-world questions don't qualify. "Will our competitor ship feature X before our release date?" will never have a public market. A watching agent works on private questions because it doesn't need a crowd.
Alerting tools fire when a keyword appears. Google Alerts tells you "someone mentioned your competitor." It doesn't tell you what that mention means for the question you're tracking. Watching agents interpret signals against a hypothesis structure and update a probability accordingly.
Monitoring dashboards track known metrics: uptime, price, volume. They answer "what is the number right now?" Watching agents answer a different kind of question: "given these signals, how likely is this outcome?" The input is not a metric feed. It is unstructured evidence from across the web.
Why this is a category
These four differences (temporal continuity, private question support, interpretive scoring, unstructured evidence) don't map to a single existing tool. You can't bolt them onto a monitoring dashboard or chain them reliably in a prompt.
The core loop: deploy an agent on a question, let it build and revise structured predictions over time. Watching Agents lets you deploy an AI agent to watch any question about the future. It builds hypotheses, tracks evidence in real time, and alerts you when things change.
The agent is the primitive here. Not the prompt, not the bet, not the keyword.
Public agents and team use
Every agent on the platform can be set to public. Public agents are browsable and embeddable, so anyone can follow the reasoning on a question without deploying their own. The growing catalog covers questions in tech, geopolitics, science, and markets.
For teams that need private agents with audit trails, Watching Agents by Inithouse has workspace tiers and a white-label option for organizations that want the prediction layer inside their own product.
If you have a question that keeps changing, try deploying an agent at watchingagents.com.
Top comments (0)