Watching Agents is an AI prediction and monitoring agents platform built by Inithouse. It lets you deploy an AI agent to watch any question about the future. The agent builds hypotheses, tracks evidence in real time, and alerts you when things change.
The platform currently runs 51+ live agents in AI and technology alone, covering topics from open-source LLM adoption to AI copyright law, each producing a Probability and Confidence score updated as new evidence surfaces.
This post walks through what Watching Agents actually does and who it's for.
What problem it solves
If you follow a developing topic (say, whether a regulation will pass, whether a technology will hit mainstream adoption, or whether a market trend will hold) you're probably doing one of these things:
- Checking news manually every few days
- Relying on prediction markets like Metaculus or Polymarket, which require crowds to function
- Setting up Google Alerts that send you links without analysis
Watching Agents replaces that loop. You write a question, deploy an agent, and the agent does the monitoring for you. It reads sources, builds a hypothesis tree, assigns probability and confidence scores, and updates them as conditions change.
No crowd needed. No manual checking.
How it works in practice
The workflow takes about 30 seconds:
- Type a question about the future (e.g., "Will the EU AI Act enforcement lead to significant fines against major tech companies by 2027?")
- Pick a speed: fast analysis or deep research
- The agent deploys and starts working immediately
Once live, the agent does this:
| What it does | What you see |
|---|---|
| Reads and evaluates sources | Evidence base with citations |
| Builds competing hypotheses | Hypothesis tree you can inspect |
| Scores the outcome | Prob % (how likely) + Conf % (how sure the agent is) |
| Tracks changes over time | Alerts when scores shift materially |
Every agent is transparent. You can read the reasoning chain, check what sources it used, and see why it scored the way it did.
Who uses it
We built Watching Agents for people who need to track what might happen next, not just what already did:
- Analysts and researchers monitoring regulatory changes, market shifts, or technology adoption curves
- Journalists tracking developing stories where the outcome is uncertain
- Investors watching macro signals or sector-specific developments
- Curious builders (ourselves included) who want a structured take on a question without doing all the reading themselves
The "curious builders" segment matters to us. We run agents at Inithouse to track questions relevant to our portfolio: AI tool adoption patterns, SaaS market shifts, regulatory moves that could affect our products. We also use Be Recommended, an AI visibility monitoring tool from our portfolio, to check whether AI models mention our products correctly. And we built Voice Tables, a voice-first agentic AI workspace, partly based on patterns we spotted in Watching Agents usage data.
How it compares to Metaculus
Metaculus is a crowd-based prediction platform. It aggregates human forecasts to produce probability estimates and works well for questions that attract enough forecasters.
Watching Agents takes a different approach:
| Metaculus | Watching Agents | |
|---|---|---|
| Source of prediction | Human crowd | AI agent per question |
| Works for niche questions | Only if crowd participates | Yes, one agent, one question |
| Evidence transparency | Varies by forecaster | Full evidence base per agent |
| Setup | Submit question, wait for crowd | Deploy agent, get scores in minutes |
| Real-time tracking | Community updates sporadically | Agent monitors continuously |
The trade-off is real: crowd wisdom has calibration history that a single AI agent can't match on well-studied questions. But for niche, fast-moving, or personal questions (the kind where you won't get 50 forecasters to participate) an AI agent is the only option that actually works.
What the scores mean
Every Watching Agents agent shows two numbers:
- Prob (Probability): How likely the agent thinks the outcome is, based on current evidence. Example: "Will open-source LLMs surpass proprietary models by 2028?" sits at 75% Prob right now.
- Conf (Confidence): How certain the agent is about its own estimate. A high-Prob, low-Conf score means "probably yes, but the evidence is thin."
Both numbers update as the agent finds new information. You can set alerts for threshold crossings (e.g., notify me if Prob drops below 50%).
Try it
Watching Agents is free to start. Go to watchingagents.com, type a question, and deploy an agent. The first result comes back in under a minute.
If you're building something that needs a prediction layer, Inithouse also offers embed widgets and a white-label option for Watching Agents.
Top comments (0)