Hi HN, we're developing Neverbell, an AI agent that combines market analysis with trade execution.

One of our beta testers said to Neverbell: “Protect my downside, but don’t be too conservative.”

Volatility increased a few days later, and the agent decreased their exposure.

Our tester wasn’t happy.

The interesting part for us was that both interpretations were reasonable.

The agent gave more weight to “protect my downside.”

The user thought that “don't be too conservative” was more important.

It changed the way we built Neverbell.

Now, if the agent sees any instructions that conflict, it will ask rather than guess what the user was trying to say.

We've seen this several times before; so we made it part of the product.

Interested in what others are doing.

Once an agent starts taking actions instead of just answering questions, where do you draw the line?

Will also answer questions about Neverbell, permissions, or anything else we've learned during the beta period.