Over the past couple of months I’ve been seeing my agent completely unable to fix bugs and just repeats the same cycle of pretending to recognize the problem, creating a fix and then coming back to me with an A-Okay message that it’s fixed but nothing at all has changed. I have bugs that I have been working on for days without any progress. And the worse part is that while I’m trying to fix one bug it creates hidden bugs in other parts of my app. I’m about to give up on Replit
4 Likes
It has actually improved quite a lot.
Have you tried high-powered mode and excruciating detail about the problem you’re facing, including where/when it happens in your UI, what the console says, and what actions you’re taking when it occurs, etc.?
Combining high-powered mode with impeccable error context usually does the trick - but it’ll cost ya.
I haven’t met a bind it hasn’t worked me out of yet.
3 Likes
I agree. The times I end up with that non-fix cycle are usually when I enter a brief prompt that “Feature X is not working”. Now I try and fully spec out in a text document what happens, what should happen, and any errors I see. I also use @Gipity-Steve method of asking the Agent to create a file with the proposed changes without actually making any changes. then I can review what it is proposing and iterate on it before giving the go-ahead. It also helps to try and break down fixes and changes into small pieces rather than 1 big change.
3 Likes
In a word @rcrownholm, no it hasn’t got worse. If anything it is much better since the latest Claude was plugged in behind the scenes.
I have mixed feeling about the high powered mode by the way. @realfunnyeric is right and wrong - sorry Eric 
Sometimes I actually get better results from the base model, and find the high powered or deep thinking models are “too clever for their own good”. They end up over-thinking and over-engineering what should be a simple solution. The resulting complex code might look and sound great in theory, because those models are tuned to be “clever”. I have the same argument about ChatGPT-5 - sometimes I want to tell it to “dial down how clever you are. Stop giving me professor grade answers to pub quiz questions”!
But, occasionally, the high powered model is needed to get around a roadblock and do some deep research on the internet about how others resolve specific issues. For example, I have been vibe coding an AI solution using Claude to do basic spelling/grammar checking on PDFs. But I have struggled to get the AI temperature right so it doesn’t hallucinate and find different typos each time I run it on the same test PDF. I dialed Replit agent up to max and asked it to research this. But immediately dial down again once I had got a good plan in my own head that I could discuss with the basic agent model.
As @CraigRT also points out, I prefer an overall strategy of “ask Agent to give me a plan and discuss changes before I give it the go ahead”, rather than “ask Agent to make changes, and then cross my fingers that it understood me and will do the right thing”. This latter model is what every vibe coder does, and whilst I know some here will disagree, is a fundamental flaw in how vibe coders are taking to this new AI dev world.
So when you discuss potential changes with Agent, you need to talk about the specifics of the planned code changes. And then finally, when you give it the go-ahead, use phrases like “ok, make these surgical changes we have discussed, and do not modify anything else”.
Finally, don’t give up on Replit - you will get there
3 Likes
Thank you all. I’ll give these suggestions a try.
1 Like
After working daily with Agent for 1 year, it has gotten better at fixing and avoiding bugs.
A couple of things I do regularly to ensure accuracy:
- Share logs Agent might not be able to see - e.g. browser console logs (F12 developer tools in browser)
- Ask Agent to review something first before starting to work on it
- Work on smaller chunks at a time
1 Like
Ok. I usually just curse at the agent but I’ll try this instead… 
3 Likes
These are all great troubleshooting tips!
2 Likes
For transparency I might’ve grumbled more than a few times at Agent 
Sometimes grit is required.
You will get it, these tips from others that have been there before should help!
2 Likes
Grit is absolutely required. And the ability to look past wasted checkpoints.
3 Likes
It definitely feels like its gotten More expensive!! Even without using the high power mode its billing has gotten way more aggressive and Im not using it any more than I was for the first couple months. Im afraid to even open it now because it seems like it takes 50$ out of my account everytime i touch it
2 Likes
Or maybe we are just starting to do more with it?
“Create me a copy of Linkedin” only cost $1, because after 3 prompts we realised it looked pretty but was never really going to fly. So we got bored and moved on to something else. Look at the Lovable scrapyard - millions of forgotten apps.
But now many of us are digging deeper and creating some genuinely proper apps with real foundations, features and revenue-generating abilities. But this costs. Of course it does, we are 1,000 prompts in to a serious app that we can deliver to real customers.
I’ve spent $2,000 on a single app. Back in my tech consultancy days I would have charged $100k to create this app for someone. The $2k is peanuts in comparison.
3 Likes
Proper take.
I’m getting great results with Replit’s no-code app, and I’ve discovered a technique that helps me overcome persistent technical challenges. When I hit a roadblock, I open a separate conversation with Claude or ChatGPT where I upload screenshots of the issues and describe what’s not working. I then ask the AI to write technical notes as if it’s briefing a developer on how to fix these problems.
I take those detailed technical notes and upload them directly into Replit’s agent. This approach has been a game-changer—it helps me communicate issues more precisely to the Replit agent using proper technical language and specific instructions, rather than relying solely on back-and-forth chat iterations. The result is clearer communication and faster problem resolution
1 Like
Gotten better for me - and - I have gotten better at understanding how to interact with my various agents (across my apps). I can’t me from a decade of my teams using Low-Code to generate apps - and - AI gen is the next wave and I will never go back.
My recommendation…
Bottom line: Treat it like a human
When I do, I get excellent results. Just like human developers it will make mistakes. My job is that of an Architect, Lead BA, and Big Thinker. For simple tasks, I just ask it to do it but try to give it as much specificity as possible (page, widget, data IDs etc).
For more complex items, I always say “don’t build yet just analyze, give me your understanding of the current context, and recommend potential solutions”.
I then get back a number of thoughtful options some of which I never thought of along with the agent’s recommendation which 75% of the time it’s the right one. This helps the agent marshal a complete understanding of the app & context of required solution to them build it or fix it well.
Sooo … sounds a lot like working with a super sharp tech lead or Sr developer to me, agree?
Hang in there, no going back!
1 Like
Getting it to analyse first is perfect, just like real developers would (should!!) do - well, those who don’t just open up a text editor and start hacking without any plan!
Btw, check out that Agent now has 2 modes: Plan, and Build.
I ran into the same problems. What you need to do is the following: Write clear instructions and present them to the Agent. 1) it is forbidden that… 2) strictly adhere to the instructions… 3) don’t add anything, change anything and focus only on my instruction and the execution of the assignment… 4) after you have submitted, you say, analyze and give feedback, do nothing until I have decided that you can get to work… 5) Make an instruction (basic) and tell the agent to save it in his core and check every morning what he made the day before. 6) Tell the agent that you want to perform a (start day check) function every day. 7) Use this command daily and that way he knows what has been processed the previous day. 8) Get a slightly higher and deeper thinking agent. Saves you a lot of time and money. 9) Think like a designer, and make sure your instructions are clear. No long-winded instructions. Then the Agent loses himself in the story.
There will still be moments that cause frustration, but, I have it pretty well under control now and still fine-tuning. Good luck Hans
1 Like
This!!! 
i actually got replit to admit it could not do something it previously had done - there is serious weakenssess in this software and it costs money to mess up - these are not apps - they are web pages - its great to build a demo but i cant see it doing anyhting else currently
agreed- i reverted to chatgpt to fix buggin issues