At Inithouse, we ship a portfolio of AI-powered products. One of them, Watching Agents, grew out of a specific frustration: keeping up with fast-moving topics by manually refreshing news feeds, prediction markets, and research papers. We wanted a tool that does the watching for you.
Watching Agents is an AI prediction and monitoring agents platform. You deploy an AI agent on any question about the future. The agent builds hypotheses, collects evidence from the open web in real time, and alerts you when things shift. Each agent produces a live Probability and Confidence score that updates as new data surfaces.
Here is how analysts and curious builders actually use it, step by step.
Pick a question worth watching
The starting point is always a question. Not a keyword, not a search query. A real question about something uncertain.
Some examples from agents already running on the platform:
- "Will the EU AI Act enforcement lead to product withdrawals before Q4 2026?"
- "Is Apple shipping on-device LLM inference in the next iOS release?"
- "Will global semiconductor supply chains normalize by mid-2027?"
The best questions are specific enough that evidence can move the needle, but broad enough that you actually care about the answer over weeks or months.
Deploy the agent
Once you type your question and hit deploy, the agent starts working. It scans sources across the open web, pulls in relevant data points, and begins building a hypothesis tree. You do not need to configure data sources or write rules. The agent figures out what matters based on the question itself.
Within minutes, you get your first Probability and Confidence reading. Probability reflects how likely the outcome seems given current evidence. Confidence reflects how much evidence the agent has found. A high-probability, low-confidence reading means "looks likely, but we don't have much to go on yet." That distinction matters more than most prediction tools acknowledge.
Watch the evidence accumulate
This is where it gets interesting for analysts. The agent keeps running. Every time it finds a new piece of evidence (a policy announcement, an earnings call mention, a research preprint, a regulatory filing), it updates the scores and logs the data point.
You are not refreshing dashboards or setting up Google Alerts for ten different keyword variations. The agent synthesizes across sources and tells you what changed and why.
We built this because we needed it ourselves. At Inithouse, we run multiple products in parallel, and tracking external signals that affect each product used to eat hours every week. Watching Agents turned that into something that runs in the background and surfaces only what matters.
Public agents as a shared resource
One feature that surprised us: public agents. When you deploy a public agent, anyone can see its question, hypothesis tree, and live scores. The platform now has a growing library of public agents covering AI policy, tech industry moves, climate targets, and geopolitical shifts.
For analysts, this means you can browse what others are tracking before deploying your own. For builders, public agents serve as a discovery layer. We noticed that public agent pages get indexed by search engines and cited by AI assistants, which creates a secondary distribution loop for the questions people care about.
Concrete use cases we have seen
Competitive intelligence. A product team deploys agents on questions about competitor launches, partnership announcements, or regulatory changes in their vertical. Instead of assigning someone to "monitor the landscape," the agents do it continuously.
Research tracking. An analyst watching AI safety developments deploys five agents on specific policy questions. Each week, the evidence logs show exactly what moved and what stayed flat. No manual literature review required.
Portfolio monitoring. We use Watching Agents internally alongside tools like Be Recommended (our AI visibility monitoring tool) and Audit Vibe Coding (our audit platform for AI-generated code). Each product watches its own competitive landscape through dedicated agents.
What we learned building it
The hardest part was not the AI inference. It was deciding what counts as evidence versus noise. Early versions of the agent treated every news mention equally, which produced noisy score swings. The current version weighs source authority and recency, which makes the Probability/Confidence curves much more stable.
We also learned that the question framing matters enormously. Vague questions ("Will AI take over?") produce vague agents. Specific, time-bound questions with observable outcomes produce agents that actually help you make decisions.
Getting started
Watching Agents is free to start. Go to watchingagents.com, type a question, deploy an agent, and watch the evidence accumulate. You can browse existing public agents first if you want to see the platform in action before committing your own question.
Inithouse builds products like this in the open. If you are interested in how we ship and measure a growing portfolio of AI tools, you can follow our work across the products we publish.
Top comments (0)