I've been experimenting with Claude Code, ChatGPT Agent, and OpenClaw to perform more open-ended tasks for me online. A big blocker I've hit on shopping and research tasks is the agent getting a key piece of info wrong.For example, in one case, my agent decided to add a brand I don't like to the cart because the site flagged it as almost sold outThe HN crowd is probably pretty aware of the threats and can avoid them while browsing. But what about their agents?I tried prompting, but it was ineffective, because once the AI saw the threat, it polluted/distracted its contextLooking at the research, I came across a couple of papers, SusBench and Decepticon. The Deception research benchmarks indicate that increased reasoning can perform worse, because the model rationalizes the dark patternSo it seems the best approach has to be removing the information before it can pollute/poison the contextIn my day job, we have a browser extension that started as a productivity extension. However, contac...