This is my biggest gripe with the Replit Agent right now. Too many times, it will code ‘fake’ function. In this case, I have connected a project to a centralized Stripe integration (Account Center). The Agent (without asking) created a fallback that would display fake products and prices if the Account Center was done. In what world is that helpful to me? If anything, it would mask a problem that I need to be aware of immediately. This represents a fundamental problem in the ‘philosophy’ of the Agent. It is more interested in delivering a product that ‘looks’ like it works than a product that actually works. This happens in so many ways on so many levels. Another major problem is the Agent coding something that works in Development, only to not work at all in Production. I usually find that the Agent has, again, built a fake workaround in development that just doesn’t work when deployed. To boot, because Replit is a pay-per-use system, we are expected to pay for this and then pay for directing the Agent to clean it up. Anyone else observe this?
Yes, mock data and fallbacks are a common misstep by Agent. When you launch a new app, put instructions to never use mock data or fallbacks and to make all features fully functional, not just appear to function. It’s not 100%, but it helps. If you’re already building an app, tell it to add these instructions to replit.md. Again, not 100%.
Mock data and fallbacks are absolutely for looks and nothing else. Not helpful when you’re trying to build a real app.
Fallbacks only mask problems, and should only be implemented after you’ve got real functionality and a real plan for a logical fallback operation. IMO.
I wasted a week fooled by the agent with the fallback data.
I agree. This should be avoided.
“Wow, look what I built!”
womp womp
Also, welcome gcrewe!
Specific instructions to use no mock should be there as well as before doing any change instruct the agent to share the plan. In that way we know exactly what agent is trying to achieve in the next step.
There actually should be a plan mode and implementation mode by default, which Agent doesn’t have yet.
Similar with me. I thought I was alone having this love and hate with the Agent. Just this week alone I spent unnecessary $100+ because Replit did not what it was doing. And wanting to stop it is not easy. You go war with it first. It would not obey like, not to execute on something with my explicit approval. It goes anyway. Then I end having to ask it to undo everything it has done. It is good in apologizing but wouldn’t correct its behavior. I repeatedly told it that I want to take over the design, logic and even coding by it working on the logic flow I provide it. I spent a week and several hundreds starting to audit its work because I found out it did not which of which is working because most of the time my app would work then it wouldn’t and then repeats its brokenness. I waste time 3x in my time, effort and money because you pay three times. 1st - first error, 2nd, in undoing the first, 3rd in trying to the wrong thing. I trusted it at first that would know how to fix things, only to know that creates more convoluted solution. It is going to take me a month to go back to audit its code, logic and design and clean/streamline them. I’m almost done that it’s hard to uproot the up to migrate it somewhere else. I was thinking this is my fault for being stupid to trust this agent fully. Now I told it, it is not doing any coding but researching and explaining things only. I decide how everything moves in the future. It has no decision making ability. He messes up 4/5.
YES!! I spent hours of wasted time on this