Is anyone else having issues with agent 4 for the past five days and over five hundred dollars worth of charges agent 4 has done nothing but destroy my development preview which I now get a 404 in and can no longer display I have to deploy every time I want to check changes and every single time I’m trying to get him to fix a certain feature that I worked so hard to get it to were it was in the first place he just seems to keep f****** it up and doing extra things that are breaking the system or doing the opposite of what I’m asking or just not doing it what he’s saying hes doing.
And every single interaction I’m having with him atm is just making things worse which is costing a ridiculous amounts of money and my charges are just going up and up and up for him to just break my system and not do anything of the kind of what he’s meant to be doing 
Agent 4 has a markedly different way of working.
My biggest observation was that the tasking system doesn’t bring context with it to the task and, upon completion, doesn’t merge context back to the main agent. At one point, I had 40 tasks going simultaneously and observed many instances where the agents took divergent paths from one another. I now have a process that lets me leverage the strength of the agents and overcome the context limitation, but it does require manual validation, and I imagine many people are struggling with the implications of this inside their projects.
Are you at a point where you can roll back to a known, stable state?
When it first launched, I spun up a few test projects to try to get a better understanding of what made it different, and I’m glad I did. It might be worthwhile.
I think it’s terrible if you have a working project, it makes things so difficult. I am switching to just using claude on my open project to be honest, its disruptive in a bad way right now.
1 Like
over 500 dollars worth of charges is insane!… Are you doing any of this in plan mode by any chance?
Oh yeah I’ve had invoices for over $700 before but replit support were really good about it and credited over half of the bill back to my account.
But no nothing in plan mode because im constantly going back and fourth with this agent trying to do the simplest of things and or even some of these things have been implemented before but now it just can’t seem to get anything right and is just constantly making things worse and I’m having to deploy every time he makes a change just to test it because it broke my development preview and can’t fix it
I went though the same thing for like 2 days. Try these things…ask your agent to always tell you what it’s doing before it does it. That helps a little if you’re not a planner like me. From everything I’ve experienced the hot load does not work on most things like 5% of the time it works so don’t feel like you’re alone! I have to publish every-time I make a change if I want to see it in expo or preview screen. Ask your agent what ports everything is pointing to. Also tell it to look for its own mistakes it makes them all the time. This also sounds a little different but tell it to tell you everything it has tried and that you need it to write it for ChatGPT then paste that. It can usually troubleshoot you to get you back going. This also made my agent find its own mistakes.
1 Like
Yeah i would use plan mode and after you ask it whatever your asking add please do a test check to make sure this functions 100%. Once i get a reply from that prompt w/e it may be i’ll either see i need to make some changes or feel confident enough to say build it and get exactly what i want. And everything @mkeene0010 said as well.
Hi, I agree with the other replies, I’d suggest two things.
1. Back up your projects to GitHub.
It gives you a second place to store your code, and the commit history makes it much easier to track down what changed when something breaks. Being able to compare commits is really helpful when debugging.
2. Use your original app plan to help troubleshoot.
Take the initial outline or build plan you started with in Replit and paste it into Claude or GPT with something like this:
“This is an app I’ve already started building, but I’m running into issues. Below is the original plan for the project. Write me an agent prompt that audits the existing codebase against the intended design, identifies likely problems or missing pieces, and suggests the most likely fixes.”
That way the agent can review the code with the intended architecture in mind, instead of guessing what the project is supposed to do. This approach has helped me many times when I’ve been stuck in a loop of trying fixes that don’t resolve the problem (which can be very frustrating).
