I’ve been writing twice a week (now thrice) about AI and agents at Reasoned, but more from a product, markets, and business decisions standpoint.
Given that Google has announced that Gemini will now act as a browser agent inside Chrome, I thought I’d share a few points on agentic browsing and delegation, based on what I’ve learned from a product standpoint while writing at Reasoned for the last month and half:
1. Browser Agents muddle the liability issue: Browser-based agents don’t just take users to a website, they act as the user: they reply to emails, submit forms, buy things, book services. To the recipient of this service, because it is you who is logged in, there is no difference between you and the browser. What you’ve done is delegate the responsibility when you actually have little control over how the browser will behave, without actually delegating the liability. This creates a new gap between responsibility and accountability in AI. What happens when the agent does something it’s not supposed to?
2. The consent tension in AI agents: The browser will ask for confirmation before performing sensitive actions. That could be a payment, but not necessarily an email. Once delegation is enabled, the power to decide when it brings you – the human – into the loop resides largely with the AI agent. Everything else happens silently. Over time, you stop supervising the agent. While consent exists on paper, it actually disappears in action. The point is that policies based on informed consent assume a level of user involvement that delegation systems are designed to remove.
3. Memory is tricky: Browser AI is moving toward continuous memory: remembering past tabs, emails, searches, and conversations to act more “personally”. This memory is a background influence, but the problem is that it has its own issues, which can cause it to make mistakes. I wrote yesterday, in “Classifieds expose the key AI fault line early”:
Context comes with its own challenges:
First, users routinely give incomplete context: There’s a huge gap between what they want and what they tell an AI model. Models fill these gaps with assumptions, and those assumptions shape outcomes.
Second, as context windows grow, models must decide what to retain and what to discard. Context selection becomes an invisible act of prioritisation, and in that process, nuance is often lost. Longer windows also trigger compression, where specificity gives way to generalisation.
Third, there is some element of what I can only describe as context pollution. Users often shift topics mid-conversation, instead of opening a fresh chat for refreshed context. Someone could be discussing automobile stocks and cars to buy in the same window, and models could mix these signals when invoking results.
Memory thus introduces a new kind of problem: not whether the system recalls, but how it forgets, reinterprets, prioritises or downgrades context over time. These are not…
Source link
Disclaimer
We strive to uphold the highest ethical standards in all of our reporting and coverage. We blogs.grocliq.com want to be transparent with our readers about any potential conflicts of interest that may arise in our work. It’s possible that some of the investors we feature may have connections to other businesses, including competitors or companies we write about. However, we want to assure our readers that this will not have any impact on the integrity or impartiality of our reporting. We are committed to delivering accurate, unbiased news and information to our audience, and we will continue to uphold our ethics and principles in all of our work. Thank you for your trust and support.
Website Upgradation is going on for any glitch kindly connect at [email protected]
