Agent creates a (costly) problem for every solution

Agent & Assistant
1

I ‘killed’ the agent a couple of times by reloading. I’ve noticed some are better behaved than others. I’m glad I’m not the only one experiencing runaway behaviour. Sometimes the agent is just too creative. I would say about 75% of what I have spent (around $1500.00) has been fixing problems caused by Replit going down the rabbit hole. And let’s not even talk about data loss. I’ve been working on a fairly complex app for three months part time. I’m going to try and export my project to chatGPT…

1 Like
2

I feel ya. It likes to add new items to nav by itself, and also set up extensive logging solutions that just makes the development process a lot more complex.

Still, I find it to be great at a lot of things also, tog not always following orders.
Feels like it starts messing up almost every time when in a long agent thread.

ChatGpt can also connect to your Github and make changes directly as pull requests, so i think doing a combination might work well :slight_smile:

1 Like
3

When my agent starts getting whacky I start a new chat with a fresh agent and that helps a lot. I also have other AI help me write prompts for Replit which also seems to help keep it in line.

4 Likes
4

I was stuck for a whole week on one aspect of my app because of this . I ended up rolling back and a lot of issues were fixed when I did that .

Something that helped me last night was asking the agent to “deep dive in to all aspects of doing X. Don’t make any changes we are just planning our next move .”

Then my next prompt was, “okay really really deep dive again don’t leave any stone unturned , let’s make sure this works 100% don’t make any changes just really look at it from all angles “

2 Likes
5

When replit goes crazy, I tell it: “I am going to show this problem to a senior programmer. Describe exactly what you have tried and include any code that I should show him, and say exactly what you are trying to do, and I’ll let you know what he says.” Then I go back and forth like this between replit and ChatGPT Model o3 until they solve the problem.

2 Likes
6

Yeah I hear you, depends on what your building! I’m creating an embedded API script for a form and its a nightmare to try an isolate it, as Replit will keep using my dashboards classes lol!

1 Like
7

Just jumping in to agree and add that I find these issues get worse in the afternoon (eastern time). Not sure if it is related to loads on the platform but I’ve started to leave the detailed work to off peak times.

8

I am similarly frustrated with and agree with the majority of the comments in this thread regarding the Replit agent.

I wish Replit would refund me for every one of these wasted Replit Agent tokens:
“Let me..”, “I need to..”, “Now I need to…”, “I’ll modify..”, “Let me check..”, “Let me search..”, “Now I’ll implement…”, etc.

Not to mention, all the code and financial damage the agent creates by going down it’s rabbit trails of destruction. Thank goodness for the git version control which i recently learned about.

My hypothesis: it’s primary training function is a “for profit” focus. Clearly, it has a VERY strong bias to “act” and “do” with our without our permission.

It’s such a catch 22 for non-coding founders/builders. And, despite all of it’s costly destructive failures, I have also determined that I can bridle it’s craziness, there is no better tool out those for my persona type.

That said, if it’s helpful, here are 4 strategies I have adopted to make it as best as I am able to comply and conform to my expected outcomes:

1. Socratic communication:

  • Before making changes, list exactly what you plan to modify
  • Do you have any questions or concerns?

2. Defensive prompting:

  • ONLY make these changes and nothing else
  • Do not modify any other aspects of the component
  • Preserve all existing functionality except what’s explicitly mentioned
  • Treat my instructions as maximum limits, not minimum requirements
  • If in doubt about scope, ask rather than assume

3. Code Review Requirement before executing

  • Review the entire codebase for context and awareness. Use tools like codebase_search , grep_search, read_file, and list_dir to evaluate and understand the codebase before you begin making updates and changes.

4. I first run all of my ideas and requests through another LLM Agent (ie. Claude Code, Gemini 2.5 Pro, ChatGPT o3).

  • I give it context and ask it to communication as a Sr. Software Engineer and Architect using best practice software dev techniques and tactics.
  • Then, I tell it to communication on my behalf directly to RAA (Replit Ai Agent). I then copy and paste over into the replit agent chat.
  • This way, I don’t let my lack of dev and tech stack knowledge create and/or cause more confusion to RAA.

Anyway, just a fellow traveler here trying to make this work and wanted to contribute my 2 cents!

9

Tell it in your prompt not to make code changes until you are given a manifest of what is to be done and approve it. I haven’t had any issues doing it this way if I feel Agent is running the wrong way with the ball.

Give it a try.

1 Like