AI Agents | Not Hacker News!

Discussion (12 comments)

Showing 12 comments

about 1 month ago

1 reply

I'm curious about the guardrails here. In my experience trying to use LLMs for user research, they tend to be "yes man" often hallucinating features or agreeing to user requests that aren't actually on the roadmap just to keep the conversation flowing.

how do you constrain the agent to stick strictly to the facts of the product hypothesis without making stuff up to please the potential customer?

Matzalar

about 1 month ago

1 reply

We ran into the same issue early on. Our fix was to lock each agent to a small JSON snapshot of the idea (no other knowledge), plus strict response templates. They can only ask questions, never describe features or promise anything. If a user asks for something outside scope, the agent replies with “not in the current hypothesis, why is that important to you?” rather than making stuff up. We also have a human review step before anything goes live.

KurSix

about 1 month ago

If you have human review for every action, then this isn't scalable software, it's consulting. Either you'll eventually remove this step for scale (and get banned by platforms), or your product will have to be very expensive (to pay for the reviewers' time). Maybe it's worth keeping AI only for the analysis part and leaving the communication to humans?

likethejade87

about 1 month ago

1 reply

Are agent pitching ideas or do actual research? Sounds super interesting though

Matzalar

about 1 month ago

They’re not pitching or selling anything they only do research. The agents ask structured questions in relevant communities and collect real reactions, pain points, objections ... No selling, no marketing language.

tene80i

about 1 month ago

1 reply

Interesting idea. Nice design. But usability issue: on mobile I hit your yellow chat CTA thinking it was submitting the app text input. You might want to move that out of the way.

Matzalar

about 1 month ago

Good point. Thanks for your suggestion. We’re very early, a lot is changing as we move on and get more feedback.

thebiggodzzila

about 1 month ago

1 reply

Chat always boosts my confidence, but reality isn’t always as kind. How can I really tell if my idea is any good beyond what Chat says, and how many people do you actually interact with?

Matzalar

about 1 month ago

1 reply

Chat is OK at making anything sound promising, but it will not talk to real people and give you the real market signals. We do both, we do the “synthetic” data analysis, but more importantly we talk to real people and ask them what they think. Depending on the product (or project), we’ll usually get few hundred real interactions. The goal is to get enough real, unfiltered feedback to see whether there’s a consistent signal to make a decision.

KurSix

about 1 month ago

The irony is that people give the most honest and unfiltered feedback when no one is asking them. When a bot approaches a user (even a research one), the observer effect kicks in, and answers often become more polite or socially desirable. Analyzing historical threads where people are just complaining to each other often yields a much more accurate and "raw" pain signal than direct questioning. There, people aren't trying to be nice to your agent.

KurSix

about 1 month ago

The idea of validation on real data is great. But why post questions directly? It's risky and intrusive. It would be much safer and more powerful to simply passively observe. Have agents scan the thousands of existing threads where people are already complaining about problems. You'll get honest data without risking being banned. People have already told you their problems, you just need to (automatically) listen, not ask again

_demo

about 1 month ago

Cool idea I have seen this people having this same problem so many times in the past.. The biggest risk I see is how do you know if people responding are providing actually usefull feedback or just garbage that will hurt me in the long run?

Resources