Tested it out today and found the UI to be very comforting/easy. I liked how simple it was to get started. My interest around wanting to test something like this had spurred on from Manus and my eagerness to try it - so this was my first take on a browser agent via prompting.
Out of personal curiosity of "what it'd take" for an executive intelligence layer to develop in the near future as AI becomes capable of retrospectively assessing the work a user has done to be able to pattern-match to 1. provide helpful insights and 2. take on the EA-level tasks of writing/capturing to whatever tool of the user's choosing, I asked Convergence to help me carry out a Deep Research task to inform me of:
1. From a technical perspective of what it'd take to process 8 hours of screen recording effectively and its cost, currently, with models such as Gemini 2.0.
2. Assess where it effectively would need to be for it to be reasonable/affordable for a user.
3. Look through historical trends in terms of price-computation to understand how it's developed up until this point.
4. Help me understand, with its best guess for how long it'd take for us to get to #2 above.
I liked how it asked me questions at the beginning which allowed me to inform it of my motive/intention and my current trust/assessed weighted believability of Ray Kurzweil's predictions & understanding of the price-performance trends.
And it went on its way.
Awesome to see, really love the idea/inevitability of it. However in practice, it completed 30 steps and went "wrong" on step 1, which is perfectly reasonable.
Upon searching Gemini 2.0 pricing, it clicked a secondary source where the headline was most similar to the literal search term it had input to begin with instead of the primary source that is the Gemini Developer API Pricing (which isn't SEO optimized comparatively lol).
And from there it gathered incorrect/incomplete information that was crucial to this entire ask to begin with.
So my assessment is that we're still early in terms of general agentic capabilities, obviously. And it needs at its core to have in the back of its "mind"/scratchpad of a weighted believability/credibility deduction system for sources, for it to really be helpful. Which, falls under the umbrella of Executive Planning capabilities, and unfortunately is a weakness even amongst frontier models like Sonnet 3.7 w/ Extended Thinking.
But on the flip-side, I ran the same test for Grok 3.0 and the response was actually like 10 - 100x better comparatively to be candid. So that's worth reflecting on as well.
Excuse the likely super weird long review in which most people don't leave, but apparently I'm on a roll as of jumping on ProductHunt yesterday, just flowing through this and having fun with it but hope it's genuinely helpful. Really like the start of it all and thank you for the experience of being able to try a general browser agent - an itch I'd been wanting to scratch for like 2 weeks now lol (now that itch is more like a painpoint in which I look forward to someone solving eventually).
TLDR: Liked the UI/UX, how easy it was. Unfortunately because everything is still in its infancy inherently, it's not at all helpful just yet. However, the fact that Grok 3.0 outperforms it (mind you, it's only 1 test though and it's anecdotal) does bring up a relevant point of functionally in terms of what tool is actually helpful, at least right now, vs. being influenced by the hype around a labelled "General Agentic" experience.
