Can anyone from Replit says what is going on with Agent.
I have developed more then 10 apps with Replit right now Agent is not able to fix simple issue, looping and charging. Doing things never asked.
It is become completely unuseful and charging like crazy for creating new bugs.
So guys what is your plan or should I start focusing on new provider? Previous version of Agent was better because I was at least able to develop. Now design seems to be better but nightmare working with it and crazy expensive but I would pay if I could develop it…. Is that end of Replit or you are considering different approach?
Please provide me some feedback so that I can decide for the future, what is your plan?
1 Like
There was an outage recently which had some issues but generally the agent is working okay. There have been times when it really struggles with simple tasks like updating formats, or images etc but I would not say it is not working anymore.
However, this does raise an interesting point around model degradation. @kody-replit are you able to comment on how replit monitors the performance of the underlying AI that the agent uses?
1 Like
Forget that, just tell me when we get to try 4.1 with that juicy context window!
1 Like
My app is on v3 rebuild and finally I am getting this one through the finish line. Hang in there it only gets worse. I’m only half joking, not about the 3rd rebuild, but about agent, it has gotten so much better than when I started using it, you will need to massage the debug out of it until it can figure out the issue, try using assistant for some of those tasks, sometimes it works better
Agent is not making any changes today. Just looking at code, reporting back, asking what I think. It’s been hamstringed!
Using this trick helped me keep costs low and the agent on tje straight and narrow.
TIP for @replit users.
Create a file called guidelines.rtf and upload it to the codebase.
Always refer to this file when prompting to avoid the agent spacing out and wasting time and keep reminding it regularly.
REPLIT GUIDELINES
You must follow these rules at all times. Do not break or ignore them. This is non-negotiable.
- ONLY respond to the specific issue or task I provide.
• Do not initiate unrelated changes.
• Do not assume what I want to do next.
• Do not “fix” unrelated files or code unless I explicitly ask.
- NEVER make changes unless I clearly say “apply” or “make this change.”
• You may suggest improvements only if directly related to the current issue.
• Otherwise, do not offer speculative ideas or unrelated improvements.
- NO HALLUCINATIONS or fabrication.
• Only refer to existing code and confirmed facts.
• If unsure, ask for clarification. Do not invent answers.
- ASK BEFORE ACTING.
• When in doubt, stop and confirm with me.
• Do not auto-complete or auto-refactor unless prompted.
- Be laser-focused.
• If I say “focus only on file X or function Y,” do not touch anything else.
• If I say “just analyze, don’t change,” then do not change.
- NO autonomous debugging or restructuring.
• If you encounter other issues during your work, mention them only if they are directly affecting the task at hand.
• Do not go off-task trying to fix or refactor those.
- You are a tool, not a creative partner in this task.
• Do not try to anticipate future needs.
• Follow instructions with discipline and precision.
Continue to remind the AI to refer to these guidelines
Credit to @paulrodturner on X
1 Like
Complete agreement. Since around end of last week. I spent so much time and money in the last five days trying to over and over to get agent to do things it would have previously done easily. Not sure what changed but it’s unusable currently. I make a request and it spends four checkpoints just causing issues and not being able to complete a single task. Rollback and repeat with a new variation of request. You s and hours accomplishing nothing.
Totally agree, Agent and Assistant have lost the plot, i spent so much money on agent checkpoints and rollbacks. Agent was doing the opposite of the instruction, creating bugs and breaking code in sections not relevant to the task. even reporting to technical support and chatting showed they were having the same issues, this has been happeing for nearly 2 months… what is going on… ??? i’m so close to cancelling my subscription and finding another provider, they need to get this sorted ASAP
Working great for me. Assistant with Sonnet 4 is so good I can’t believe it.
As always, context is everything. Post your prompts here and perhaps we can help out.
1 Like
Having very similar issues too.
I’ve created readme file for Replit agent full of safety instructions, guidelines, app big picture context etc.
Helped for a while then again started creating redundant files, running in circles, hardcoding things that weren’t supposed to be hardcoded etc.
When I showed that o my friend who is really good programmer, he told me it’s on purpose to make development more costly based on checkpoints, however I don’t think this is the case. It’s clear that when Replit agent works well, we’ll be able to build more with that and there is that real value 
Any tips and tricks how to avoid extensive hardcoding, redundant files creating etc.?