I have spent 4 days on a simple weekly sales calculation card and the agent can’t figure out how to get it to calculate properly, then can’t get the display to work. I asked it for a very simple data value extraction and it added a new tab and created 3 test buttons. Can we Roll back to old agents please?
i have turned off experimental features in my replit profile, and I’m so much happier, and wish I’d done that 2 weeks ago.
Really? What have you noticed?
I’ve no idea if it’s using claude 3.5 or 3.7, but i do find that it just does what i tell it, and pretty much stays on track during the same chat, rather than wandering off and doing random stuff, or undoing what it did earlier. It feels like agent 2 is managing the context window differently and losing some of it over time, which agent 1 somehow got right (a tough ask). Point is, I did a LOAD of dev in Jan/Feb, and came to a halt in March with very little progress. I thought it was my app getting complex, but no, it was this agent change. I’ll be very cautious with experimental features from now. The $100 spend was manageable, but my lost time is irrecoverable.
3 Likes
You nailed it exactly, I got more done in my February than all of March due to the constant need to force the agent to focus. I wish I had worked faster and knew more I would have had my app done a month ago. Shame, it was almost done. Just can’t get the agent to put that last brick on without smashing through the other walls.
1 Like
Wow. I’m going to revert today. Did you have any issues switching mid project?
no. but make no mistake, agent v1 is not what it used ot be. i think the system instrucitons have changed to suite agent v2. but it is definitely better. Wish i could get a credit for the $100 i wasted this past month, and the $10’000 of my time!
1 Like
I turned off explorer mode and had a great weekend. I got more done over the weekend then I did last in the 3 weeks. Monday hits and seems I am back at the slog. I am not sure if they changed the algorithm again but now the agent is back to doing a lot of actions with no result. Maybe the user load was so low over the weekend I was able to use more computing cycles?
I think you get v2 irrespective of that setting now. Having said that, I found from around Sunday, it has actually been rather good. My general rule now:
- New pages, give to agent. It does stuff I didn’t even thing of.
- Changes, stuck to Assistant. It’s like Agent has a high temp and top k, so get too creative, whereas assosiate does the job more cleanly, but only a few steps at a time
1 Like