My top recommendation when it gets in a loop, or you just feel the vibe has gone and it no longer knows what it is doing, is to start a new chat.
You have to give the new chat a bit of context to bring it up to speed. But then give it the problem again, and chances are it will look at it with a fresh head rather than loop round the same bad decisions.
my take: Replit should carefully evaluate its cost model, especially with more affordable alternatives emerging. A drastic change would be unfortunate. By the way i was not allowed to type the word s h i f t
It is really sad how they have increased this pricing similar to extortion now that they know we are using it.
I wrote to this Replit support , cos when i comp,mained about AI making mistakes and I paying for those mistakes, they said āits Just how the AI behavesā. This is what I wrote back. 'Now every time i engage with the agent I get hit by ridiculous pricing and I have to pay for AI mistakes over and over again.
Looking at the communityās reaction I also see people are reacting negatively to this.
Can you please let your decision makers know that this isnāt fair. Earlier in your reply you said the mistakes are just how the AI works, then if you know this price should be reflective of this.
It is like you ordering a chicken dinner at a restaurant and instead you get beef ribs and when I complain you say ā oh yea the chef does this sometimesā, but you already charged me and thereās no refund. Will you ever return to my restaurantā??.
I love Vibe coding because I love to be in the flow. If the app destroys my flow I go. Extensive costs and Agent that overcomplicates simple things is what put me out of the flow.
Most platforms think they are special with adding new features. Not at all, you are special if your platform keep me in the flow so that I have great time using Platform.
Is that simple.
The princing for AI assistant and Agent and agent is becoming ridiculous in replit. It becomes really expensive when the agent make mistakes and you have to make another promt to fix mistakes.
Assistant hasnāt changed in price since Iāve been a user.
Agent w/out $ or $$$ activated features (extended thinking or Opus 4) is largely the same, it just doesnāt produce five 25-cent checkpoints, annoyingly, and now will just return a single response with effort based pricing around $1-$1.50
In general, my experience has been pretty positive given that we are dealing with relatively new technology operating is a very new and uncharted use case. These are probabilistic models so they will lose their way and make mistakes just like humans do. Iām not a coder so my options are paying someone else hundreds or thousands of dollars to develop an app for my clients. Or, I can work on it myself, learn something along the way and own the IP for a few hundred dollars. If I encounter a problem that is critical to my app working and it costs me $5 to fix it, Iāll sacrifice the cost of a latte and take the win.
I am starting to run into what, to me, seem should be straightforward fixes. No new features, no change to logic. Just an error during workflow. What I am finding is that no matter whether I use assistant or Agent then I get a very enthusiastic response that it sees the issue and can fix it. Only for the exact issue to remain. I do that same process a few times with no fix. I then go to Agent and the same happens. Then add extended thinking and the same. Finally full on extended thinking, high power model, long prompt asking it to re-engineer the feature if needed. Sometimes even a few attempts. Finally a fix. If I was dealing with a consultant at this stage I would be insisting that they now fix it for free since they had cost me a lot of time and money after already saying they knew what the problem was and how to fix it. But there is no recourse with an AI.
Welcome to replit!
A cynic would say this āfeatureā had been designed by their finance team 
I struggle with this from time to time, especially when I have similar functionalities in different places. It makes the repair to the wrong modal or area, so I think it isnāt working, but it is, just not where Iām looking.
Get very specific about file names in your prompting, this sometimes helps.
Iām aware this may not be applicable to your situation, but itās a trap Iāve fallen into a few times.
I usually prompt using the page, my action and the error rather than file names as I donāt know in which file he problem might be. If there is a crash I past in all of the error text. Iām not sure what else I can prompt.
Thatās a good workflow. But if itās saying itās made changes and youāre not seeing them at all - for me at least - itās usually working in the wrong place.
I sometimes say āshow me the before and after code for what you changed, together with the file names/foldersā.
But as Iāve said before, my preferred approach to vibe building is: ātell me what you want to do, and I will give you the go-aheadā. So based on this, I often ask agent to:
āshow me the before and after code for your proposed changes, together with the file names/folders. We will then discuss it before I confirm you can proceedā.
It leaves Agent in no doubt as to who the big boss is
If I could guarantee the result I want on the first try, Iād sometimes be willing to pay $200/prompt on some of the heavier, difficult stuff that takes me days. Just depends if I was guaranteed it would work or not 
Thatās actually an interesting thought. Make an AI ābountyā on a given feature. Make it to where we can outline and very specific request and if the request is successful we would be willing to pay X dollars for it. Then the AI can scale up or down its power as needed to complete the task within the budget. A lower budget bounty would take longer whereas a high value budget would be more resource intensive but ultimately more successful quicker.
I wish it was so simple. This past week I seem to be going through an experience where the higher level models are utterly useless, and when I switch back down to the basic model it works like a dream and understands everything I ask.
Next week, I am sure the experience will change. Nothing like AI dev to keep us on our toes! 
Beauty is sometimes in the simplicity
Itās definitely helpful to change them around. I also will use a mix of Agent and Assistant, or plug code into external AIs when I hit a roadblock to review whatās going on.
@Gipity-Steve I agree with you. My cost for the same model has now at least tripled after switching from the .25 prompt to the per-prompt one. None of my prompts cost under .80 today. And that five bucks for the superior model isnāt worth it; itās better spent asking the standard model five times rather than asking that one once. Itās not worth five dollars. Not to mention, Replit keeps removing features, making our lives harder instead of easier. Dev tools are back after our complaints, but theyāve removed the visual editorās direct navigation to the selected container, which probably makes manual debugging harder.
Iām perplexed as well.
One prompt cost me 50% of my agent allowance that wasnāt my error - the error was created by the agent itself. I even prompted that I needed to input my own HTML (prompt below) yet it went ahead and created everything including all the mermaid tabs I never asked for, without my code, but with the error that took it 4 minutes to fix:
Below shows how long it took to fix the error that depleted my agent allowance by 50%.
Here was the agentās self-generated server error:
