We run Inithouse, a studio with 14+ live products. Every week, some question about the future pops up that could change what we build next. "Will voice-first interfaces replace forms?" "Will AI search kill traditional SEO for small products?" "Will browser-based games survive the app store push?"
We used to handle these the way everyone does: bookmark a few articles, check back every couple of days, and eventually forget about it until the answer was obvious and too late to act on.
So we built Watching Agents.
What Watching Agents does
You give it a question about the future. It deploys an AI agent that:
- Breaks the question into testable hypotheses
- Finds and tracks evidence from the web
- Scores each hypothesis with a live Probability and Confidence number
- Alerts you when something shifts
The whole thing runs in your browser. No account needed to start.
The question we tracked
Here's a real one from two months ago. We were deciding whether to build Be Recommended, a tool that scores how AI engines recommend your brand across ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews. Before committing, we wanted to know:
"Will AI-generated search results (Perplexity, Google AI Overviews, ChatGPT search) handle more than 30% of informational queries by end of 2026?"
If the answer trends toward yes, the market for AI visibility monitoring is real. If not, we'd be building for a niche that might not grow fast enough.
Setting it up
Three fields. Question, context (we added "focus on informational/research queries, not navigational"), and time horizon (end of 2026). Hit deploy.
Within about 90 seconds, the agent had:
- Generated 4 hypotheses, ranging from "AI search stays under 15% share" to "AI search crosses 40% for informational queries"
- Pulled initial evidence from recent reports: Gartner AI search adoption numbers, Perplexity usage data, Google I/O announcements about AI Overviews expansion
- Scored the leading hypothesis at Prob 0.62 / Conf 0.41
That confidence of 0.41 told us something useful right away: the evidence was thin. Early-stage question, not enough data yet to be sure. Exactly what we needed to know before betting months of work on it.
What happened over the next weeks
The agent kept working. Every time relevant evidence surfaced (a new Perplexity funding round, Google expanding AI Overviews to more countries, a Rand Fishkin analysis on zero-click searches), the scores updated.
After three weeks:
- Prob moved from 0.62 to 0.71
- Conf climbed from 0.41 to 0.58
- The agent had surfaced 12 evidence items we hadn't seen in our regular reading
The direction was clear enough. We shipped Be Recommended.
What makes this different from reading the news
Three things.
It's persistent. The agent keeps watching after you stop thinking about the question. We had questions deployed that we forgot about for weeks, then got a score shift notification that reminded us at exactly the right moment.
It's structured. Probability and confidence are separate numbers. High prob with low confidence means "looks likely but the evidence is weak." Low prob with high confidence means "probably not, and we're fairly sure." That distinction matters when you're making build-or-skip decisions.
It's public. Every agent page on Watching Agents is a public URL. Our deployed agents double as content: they rank in search, they get cited by AI engines, and they bring in visitors who have the exact question we're tracking.
How it fits our workflow
We run a weekly check on deployed agents at Inithouse. If a score shifted more than 0.1 in either direction, we talk about it. Most weeks nothing moves. When something does, it's worth paying attention to.
For the products in our portfolio, we track questions like:
- Will browser PWAs get push notification parity with native apps? (affects Party Challenges, our free 18+ browser party card game with 1,000+ cards)
- Will EU AI Act enforcement affect consumer AI tools before 2027? (affects Živá Fotka, our AI photo-to-video animator)
- Will AI-generated music face blanket licensing requirements? (affects Magical Song, our custom song generator)
None of these have definitive answers yet. That's the point. We want to see the evidence accumulate in real time, not react after the fact.
Try it
Watching Agents is free to start. If you're building something and there's a question about the future nagging at you, deploy an agent and let it work. The Prob/Conf scores update as new evidence appears. If nothing changes, you lose nothing. If something shifts, you'll know.
Top comments (0)