Ideally, Replit Support or the Agent should function like any normal app you spin up directly on Replit. But right now, it feels like they built the so-called v3 Agent using v2 under the hood—and the result is essentially a crappy AI wrapper.
The Agent struggles with even the most basic tasks (like setting up a backend) and the cost is insanely high. At one point, it charged me $51 in a single session when it got stuck in a loop—constantly calling architecture routines or endlessly searching the codebase.
What’s worse, I’ve even seen the Agent randomly start on its own, ruin work in progress, and still rack up charges. Instead of fixing this, support blamed me for “not monitoring my usage,” which is completely unfair given that these failures are clearly systemi
it was ours fault trusting a random so called app replit , literally burning money with doing nothing , trying using Claude code directly so replit markup and poor performance is eliminated , install Claude code directly in your shell
This is so depressing. I’m in the process of moving my code off just in case, it will delay me months now so I’m hoping they just revert the changes or offer extended use of agent 2 for us with older code. There is no business use case for this system if they are making decisions like this, this cannot be mission critical in any way shape or form with this type of irresponsible rollout. Did no one beta test this? Was anyone asked to try it?
your code is corrupted on their server which was v2 , so not possible to continue further , basically you have to rebuild and spent same time and money with their new agent to continue
Let’s stop sugarcoating this and this is already too much with those interview and marketing spend over v3 and it turns out to be a joke : Agent v3 is a trainwreck. It’s not a developer tool, it’s a glorified AI-powered loop generator wrapped in buzzwords and sold as innovation. Anyone actually building serious applications can see this thing is architected like duct tape on duct tape.
The issues are not bugs, they’re fundamental design failures:
No persistent memory → the Agent can’t even hold context for more than a few steps; try working with a 4000+ file repo and watch it collapse.
Execution reliability = zero → audits fail, backend setup is impossible, loops spawn endlessly until your wallet bleeds out.
Ghost processes → the Agent literally launches itself and trashes code while burning through credits. That’s not “AI,” that’s a rogue process masquerading as a feature.
Billing black hole → one event cost me $51 for a stuck loop, and when $4k+ goes down the drain, your support team blames “user monitoring.” That’s professional gaslighting—it’s your unstable orchestration engine failing, not my oversight.
And yet, all we hear from Replit’s leadership is glossy, performative hype about “AI-native development” and “agents changing coding forever.” No—they’re not. Right now this is an overpriced, underbuilt, half-functioning prototype being marketed as the future.
Let me be clear:
If your Agent can’t persist state, traverse code without choking, or avoid recursive self-destruction, then you have no right calling it a “developer tool.” Developers don’t need trash bots that hallucinate architectures and light money on fire—we need stable systems that earn trust instead of eroding it at 100x cost multipliers.
Replit devs: this isn’t shipping innovation. It’s shipping liability. And every time you push PR spin instead of fixing the orchestration core, you’re proving exactly how disconnected you are from real engineers trying to use this in production.
Cut the interviews, cut the hype. Fix the foundation. Right now, you’re bleeding user trust faster than your Agent burns through compute loops
I tried to build something into one of my projects to protect my own codebase from the agent…
An anti-drift mechanism!
But alas, there’s no way to really enforce this w/o Replit really handling.
When starting a new chat i tell the agent to reference my rule file. It explains the project, files etc. THEN it also reminds the agent to only change what is needed.
THEN it tells the agent to create a ‘ticket’ as part of the plan before it builds.
The ticket targets certain files for changes, lists them, approximates a budget for line qty needed.
The initial rule script also creates checksums on the project files. Then after the agent finishes its tasks it’s supposed to run a checksum script - which will point out if files outside of the scope were changes. And if so, we can assess if other areas were borked.
Anyway - it kind of worked. But it wasn’t 100% running the scripts, and really required me to have a heavy hand-hold.
Doesn’t address core issues, but might help put guard-rails or bring transparency in.
i’m trying out just using the Assistant again like the old days.
Slower, but using the learnings of ‘planning’ then ‘building’ is helping.
For example - i have one chat for planning and one for building.
prompt something like ‘build a plan in file buildplans/working/xyz.md to do xyz, and include acceptance criteria’
Then it creates the file, i move to other chat and say ‘inspect buildplans/working/xyz.md and ask any questions you may have before proceeding’. then it asks me questions, or not, and then it does work which i manually click ‘apply’.
Def more hand-holdy. But WAY cheaper, and honestly it’s not borking stff up.
I’ve only ran this today though…so if anyone else wants to try and let me know. Claude4.0 btw
Excellent post. A week ago I was in two minds as to whether to wait and see if Replit actually tried to fix this or just rip the band aid off and move. I did the latter - to Firebase. Glad I did.
Even if Replit either (a) fix this and/or (b) allow you to select Agent 2 instead, there is ZERO trust now that in the future (A) they will not break this again and/or (B) remove the option of Agent 2. Replit self evidently has total contempt for its user base and just sees them as economic units to be plundered. I have abandoned ship - drowning would be preferable to being screwed from behind as the ship slips under the water.
What do you find so terrible with it? Admitedly it is struggling to get the environment configured correctly so I can actually view the app! Got it going once, now I am in port management hell
Really? I find the interfaces almost identical. The issue for me is that replit sorts out all the back end stuff well. But Firebase seems pretty convoluted. I imagine if I started from scratch within firestore it would be ok, but imports seem to be a bit of a mare.
First, open the Shell in your Repl (if it’s not already open)
2. Run these commands in sequence:
git fetch origin git status git pull origin main
Please share the complete output from these commands. This will help us understand exactly what’s causing the error and determine the best way to resolve it.
Also, before you proceed:
Do you have any uncommitted changes in your current branch?
Are there any merge conflicts shown in the Git pane?