Most market questions have a shelf life. You Google something in March, bookmark three articles, and by June the landscape has shifted. The bookmark is stale but you don't know it yet.
At Inithouse we built Watching Agents to fix exactly this. You give it a question about the future, set a time horizon, and an AI agent watches it for you. Not once, but continuously, until the question resolves.
Here is what that looks like in practice.
Pick a question that matters to you
The starting point is a question with a time horizon. Not "what is the best JavaScript framework" (that is an opinion) but something like:
- "Will the EU AI Act trigger its first major fine by Q2 2027?"
- "Will Apple ship a dedicated AI device by end of 2026?"
- "Will Rust overtake Go in TIOBE top-10 by 2028?"
These are questions where the answer changes as new evidence appears. You could track them manually, but you probably will not. I know I stopped checking my "Will WebAssembly replace containers?" question about four months in.
Deploy an agent on it
On Watching Agents, you type the question, set the resolution date, and deploy. The agent starts immediately. Within the first cycle it does three things:
Builds initial hypotheses. For and against. Not vague "maybe yes, maybe no" but structured reasoning: what would need to be true for each outcome.
Finds starting evidence. Sources, signals, data points that exist right now. Each piece of evidence is cited and linked.
Sets an initial probability. A number, not a vibe. The agent commits to a Prob/Conf score and you can see exactly why.
This first snapshot is already useful. But the real value comes from what happens next.
The agent keeps watching
Every time relevant new evidence appears, the agent picks it up. A regulatory announcement, a product launch, a quarterly earnings mention, a research paper. The probability updates. The hypotheses get refined or discarded.
You get alerts when something meaningful changes. Not daily spam, but actual shifts in the evidence base.
Here is what makes this different from setting a Google Alert: a Google Alert tells you a keyword appeared somewhere. The agent tells you whether that appearance actually changes the probability of your question resolving one way or another. Context, not noise.
What a probability timeline looks like
After a few weeks, you have a versioned history. The probability started at 35%, jumped to 52% when a key regulation draft leaked, dropped to 41% when an industry group pushed back, climbed to 58% after an enforcement precedent in a different jurisdiction.
Each update has a timestamp, the evidence that triggered it, and the reasoning. You can audit the whole chain. If you disagree with the agent's reasoning at any point, you can see exactly where your model diverges.
We use this internally at Inithouse to track questions about our own market. "Will voice-first interfaces reach 10% of enterprise data entry by 2027?" is one we are watching right now. The probability has bounced between 12% and 28% over the past two months as new voice AI products launched and usage data trickled in.
When the horizon arrives
When the resolution date hits, the agent does a final assessment. Did the thing happen or not? The full evidence trail is preserved. You can look back at every hypothesis, every update, every signal that mattered or did not.
This is useful for calibration. You can see where your intuitions were right, where the agent caught something you missed, and where both of you were wrong.
Public agents as a starting point
Watching Agents also runs public topic-level agents on broad domains: technology, economics, geopolitics, science, and others. These are open for anyone to read and follow. If you want to see what the output looks like before deploying your own agent, start there.
The public agents cover questions submitted by the community, so the topics tend toward whatever people actually care about rather than what an editorial team thinks is important.
Try it
If you have a question about the future that you keep Googling every few weeks, that is exactly the use case. Deploy an agent on watchingagents.com and let it do the watching. Free to start, no credit card required.
The best use I have found so far: questions where I have a strong opinion but no systematic way to track whether reality agrees with me. Turns out reality often does not, and finding that out three months earlier is worth something.
Top comments (0)