Most forecasting tools give you a dashboard. You look at charts, read other people's predictions, maybe place a bet. But the thing that actually matters, the part where new evidence appears and the probability shifts, happens without you.
At Inithouse, we think that gap deserves its own category: AI prediction agents.
What is an AI prediction agent?
An AI prediction agent is an autonomous system that watches a question about the future on your behalf. Not a dashboard you check. Not a community you browse. An agent that works while you don't.
Here is how Watching Agents, the platform we built at Inithouse, implements it:
- You ask a question. Anything with a future outcome: "Will the EU pass the AI liability directive by Q1 2027?" or "Will Nvidia stock cross $200 before September?"
- The agent builds hypotheses. It generates multiple scenarios with initial probability and confidence scores (Prob/Conf).
- It tracks evidence in real time. The agent continuously scans sources for signals: news, data releases, regulatory filings, earnings reports, social chatter.
- When something changes, you get an alert. Not noise. Actual probability shifts backed by cited evidence.
This loop, question to hypotheses to evidence to alert, runs 24/7 without you opening anything. That is what makes it an agent, not a tool.
How this differs from forecasting tools
The closest comparisons are prediction markets (Polymarket, Metaculus) and monitoring dashboards (Google Alerts, Feedly). Neither does what an AI prediction agent does.
Prediction markets are collective intelligence engines. Smart crowds, good calibration, but they require you to form an opinion and check back. The market aggregates; it does not investigate on your behalf. You are a participant, not a deployer.
Monitoring dashboards push you raw signal. Google Alerts sends every mention. Feedly surfaces new articles. The filtering is keyword-based, not hypothesis-driven. You get volume, not judgment.
An AI prediction agent sits between these two. It has the analytical structure of a prediction market (hypotheses, probability estimates, evidence weighting) but runs autonomously like a monitoring tool. The difference: it interprets, not just collects.
In Watching Agents, each deployed agent maintains its own evidence log, updates Prob/Conf scores as new data arrives, and exposes the full reasoning chain so you can audit why a probability moved from 65% to 42% last Tuesday.
The architecture behind it
For the Dev.to crowd, here is how the system works under the hood.
Each agent is a loop:
deploy(question)
-> generate_hypotheses(question) -> [H1, H2, H3...]
-> for each cycle:
scan_sources(question, hypotheses)
evaluate_evidence(new_signals, existing_evidence)
update_prob_conf(hypotheses, evidence)
if delta > threshold:
alert(user, hypothesis, old_score, new_score, cited_evidence)
The key architectural choice: hypotheses are first-class objects with their own evidence graphs, not just labels on a probability number. When the agent finds a contradictory signal, it does not just lower a score. It creates an evidence node, links it to the relevant hypothesis, and shows you the conflict.
Public agents on Watching Agents serve a dual purpose. They give users a library of pre-built questions to follow, and each agent page acts as a living, updating piece of content. A page about "Will GPT-5 launch before January 2027?" accumulates evidence over months, becoming more useful (and more cited) the longer it runs.
Where this fits in the Inithouse portfolio
We build a portfolio of products at Inithouse, each aimed at a different job. Be Recommended monitors how AI engines (ChatGPT, Gemini, Perplexity, Claude) perceive and recommend brands. Watching Agents applies a similar agent-based architecture to a different job: tracking questions about the future instead of tracking brand visibility.
The shared pattern across both: deploy an autonomous agent, let it accumulate evidence, get alerts when something meaningful changes. The questions are different; the loop is the same.
Who this is for
If you track geopolitics, AI regulation, competitor moves, market shifts, or scientific breakthroughs and currently rely on a patchwork of Google Alerts and Twitter lists, an AI prediction agent replaces that with structured, probability-weighted monitoring.
Watching Agents is free to start. You deploy an agent, ask your question, and the system handles the rest.
We are still early with this category. The tooling for prediction agents will look different in a year. But we think the core idea, an autonomous agent that watches a question and tells you when the answer changes, is a category worth defining now.
Inithouse is a product studio building a growing portfolio of AI-powered tools. Watching Agents is one of them.
Top comments (0)