I would definitely +1 the idea of Replit forming a testing group of real customers. Both experienced devs and complete novices.
I’m not sure of the details, but like many tech companies, I sometimes worry their engineers only ever talk to each other, in a windowless basement room, with just a whiteboard for a friend.
It is time they got outside and included real customers in their product development.
If they’d done that with Agent v3, we would have headed them off at the pass long before it ever embarrassingly got out into the real world!
1 Like
Yes, so the big question is, how do we manage this? Who’s going to put the turd on the table?
1 Like
When I spoke to @FranciscoCM, he said they are looking at various ideas to bring the community closer.
I am sure I recall a user test group being mentioned. Or was it just my wishful thinking Francisco? 
2 Likes
That would be more than great!
1 Like
I can assure you this isn’t the case. And agent 3 seems pretty close to agent 2 if you pick the right settings now. You can go nuts if you want to, but you can also dial it back in. My workflow is back to uninterrupted as it was pre v3.
2 Likes
Hey @Gipity-Steve 
We have kicked off a very small cohort to begin setting up the infrastructure for a much larger group in the future! I am aiming to iron out some of the workflows with the internal team and then launch an invite only program early next year before a wider release soon after.
Mainly will be looking for a diverse set of users that are experimenting with the tool in different ways and also have a track record of being helpful within the Discourse communit🙏
More information to come soon!
2 Likes
I’ve been working with the “medium” version for a few days now. Aside from the fact that it still sucks, they somehow managed to send two invoices within two days (2x €60.60). What the heck!
It’s only getting worse. Yesterday I literally spent 6 hours on a simple correction with a clear prompt. Adjusting a widget. Text changes.
Constantly working in files that are in quarantine. Several times you get back: “sorry, I was wrong.” Zero solutions…
It really is becoming a problem. What bothers me most is that there are no refunds whatsoever. In the end, we’re the “developers” who add value to Replit. Based on that, Replit develops!!! We’re part of a test environment and have to pay for it. I find that really outrageous, to put it mildly.
I really wish we’d all make a stand here and ask Steve if you’d be willing to put time into this again. (You’ve spoken with the guys before.)
I’m slowly getting fed up.
1 Like
I’m sorry you’re having trouble. Unfortunately (well, fortunately for me), Replit has been singing for me, both on new apps and on apps previously created with v2. I’m experiencing costs that are more in line with v2 as well.
Can you explain more about your app, it’s size and function, and perhaps we can drill down at the root cause.
2 Likes
I’ll definitely keep nudging on here, but otherwise it sounds like @FranciscoCM has got a great plan for a bigger program - so we just need to watch and wait on that.
On “refunds”: I can see how this is probably impossible to implement and police - we all know there will be people in the world who’d abuse it. Also, there is a certain element (not necessarily yours) where it is “user error”. Like taking a wrong turn when driving our car, which causes an extra 100 mile detour, and then asking the petrol company to reimburse us. If you can think of a model that would work then let us know.
Agent doing stupid stuff: Can I ask one thing, how old/long is your current agent chat? Sometimes a very simple solution is to start a new chat, because the old one becomes tired and groggy - its context window has been blown to bits and it has lost track of everything. You could ask it what 2+2 was and it will get it wrong. Only answer is start a new chat. Of course, your issue may be completely different. But if the chat is old or long, I always start there.
1 Like
Hi, I can’t tell you anything about it. I’m bound by a Non-Disclosure. What I can say is that it will be a platform rolled out in the Benelux and later across Europe.
My pain point is that the AGENT sometimes takes a wrong turn, despite all the do’s and don’ts. Short prompts and clear instructions. And above all: what is forbidden.
My database is organized, I program in HTML and export to Wordpress Desk with a separate flow for WPA. And that’s where the bottleneck and frustration come in. The design differs between Desk and Mobile. That makes it costly and requires extra thinking. The foundation is perfectly fine, and agreed upon.
But what happens, despite all the instructions and forbidden things, is that the Agent—while editing in the mobile function—also applies those changes in Desk. That’s really awful.
So, besides the desk function, I also have to create the mobile function, which deviates from the PWA.
You understand the complexity.
I work in LOW and Medium, and sometimes I’m at my wits’ end. How then? What works, what doesn’t?
I’m not even talking about the costs, because those are getting out of hand as well.
If you have any tips, I’d love to hear them.
I hope to be able to show you more soon.
“Agent doing stupid things: may I ask you something, how old/long is your current agent chat? Sometimes a very simple solution is just to start a new chat, because the old one has become tired and groggy.”
Yes, sometimes I stay in the same chat for a whole day. I’ll try starting a new one. (learning curve)
1 Like
Tell me more, how did you do this?
I can often “feel” when the agent is getting tired. As soon as it does something that makes me think “whooah, that was a weird/odd/stupid thing to do”, then it is time for the chat to go to the big AI heaven. And start a new chat.
But here’s another thing: when it does that weird thing, I also ask myself:
- have I been vibe coding for a few hours straight without a break?
If the answer is yes, then sometimes the problem is me! I am tired and my prompts are probably getting sloppy and less focused. So time for a cup of tea and piece of cake - away from the computer*
* or whatever else you people in Amsterdam prefer to do to relax 
1 Like
I never go to Amsterdam. Too busy, too many tourists, too many drugs, and too expensive.
I’m from Utrecht myself, the nicest city in the Netherlands.
But you’re right, I spend too long on one chat, at the end of the day you just trust the Agent to understand what you’re doing… The biggest mistake you can make.
It also drags you along in your own flow and enthusiasm, until the bills come back around… 
1 Like
Sounds like with this app you may need to stay in Low mode all the time otherwise the architect agent is going to come in and try to outsmart you.
You may also want to just use Assistant for surgical precision. It’ll take longer but won’t deviate from explicit instructions.
Both of these options are still capable of breaking things, especially assistant, but with less deviation.
1 Like