I also found this out:
Yes, the joys of AI. I have had several instances of spending a lot of time and effort (and therefore $$) to “fix” these Agent created problems. As I have said elsewhere - if a consultant or agency did this you would be asking for money back or the fix be free but with AI you have no recourse but to chalk it up to the cost of doing business. 
The first prompt of my first app worked way better. The difference was I knew exactly what value and exact feature(s) I wanted - telling it what to do ‘at that moment’ and not ‘what I’m planning.’
From the logs, the agent was simply doing what it was supposed to do, and it encountered port conflicts, which can arise from a reboot sequence. To address this, it added a workaround.
I see a cost of 58 cents - was there something previously that caused a 50% depletion in credits?
Thanks for the reply. No - the 50% depletion occurred right after it self-automated resolving it’s own server error in the 2nd app (Mermaid).
1 Like
What was the cost of fixing it? Did you use the high powered mode ($$$)?
I didn’t change anything.
@Gipity-Steve Well put & 100% agreed - I experienced exactly the same.
The Agent made more mistakes, took more iterations to achieve what I know would have been more easily done in other sessions. Throttling lower tiers to sell higher tiers (following Apple and Openai steps).
- I tried extended thinking and found that it was obviously costlier with similar results. After turning it off, it seemed everything was extended.
- There were builds that took 30-45 minutes with hefty price tags.
- Granted, more work = more cost, BUT this is work that it handed easily before.
- So either the Agent was lesser or the effort was priced higher,
- @andrewjdavison asks a great question: “Was it worth it?”
- Not necessarily, because the mark is often missed, requiring more iterations
1 Like
Sad to say I’ve paused all further development on Replit and I’m only doing basic maintenance until the pricing issue is fixed. Hope they find a solution soon and give us options, as my work is accumulating. Hate to go look for alternatives this far into the game.
2 Likes
I’m in the same spot. If they don’t find a solution, I need to rethink my approach and maybe even my full use.
It’s a slippery, costly slope.
1 Like
Is anyone else having trouble with their agent getting stuck at the login screen during debugging? Before Agent 3, my agent could test within the login screen without any issues, but now it can’t even get past the login step, and I get billed for the time it struggles to log in.
And in frustration, I was billed $0.47 for requesting a button to toggle on “Agent2“ 
This is especially true with a front-end and back-end setup.
I’ve found it helpful to define specific areas of the system when working with the agent, then reference those areas while requesting changes.
For example, I’ll say, “In the admin UI…” Sometimes, the agent still makes mistakes, so I’ll ask, “Are you sure you made those changes in the admin UI?” It then double-checks.
I’m starting to think Replit deliberately introduces errors just to elicit a few more prompts from me.
For anyone else who, like me, is about ready to throw in the towel: Disabling the “App Testing” mode and setting the autonomy level to “Low” has made the agent feel and behave much more like it did before Agent3.
I wasted hours and dollars farting around last week with Agent3, and getting nowhere.
1 Like
I can now say that nearly every prompt I make is $5 or more. 
1 Like
Life comes at you fast!
1 Like
On what autonomy mode? And what kind of prompts?
I do plenty on low/medium/high that cost under a dollar - many 10-50 cents.
Mostly medium and some high. I do have some that are around that cost, but not often.
I’ve been going through and cleaning up a lot of Agent 2 code, unit and user testing, security/bug fixes, and my prompts have a lot of work for Agent, so it’s running for like 30m-1h15m each prompt in most cases.
I’m definitely not complaining about it, I thought my release timeframe would be ~ November, but in the past 48-hours Agent has been awesome and nailing everything I’m giving it.
1 Like
I think I’ve said enough elsewhere about my thoughts on this. Maybe my OCD and control-freakery is coming out, but it frightens the living daylights out of me, allowing the agent to do that
Oh, believe me, it will scare me too once I’m in prod. Right now though it’s not high risk stuff and could roll back quickly if it breaks stuff.
1 Like
To the owners of Replit.
I’ve already paid quite a few invoices, but I’m starting to get a bit worried about the future. It can’t be that building a project/platform will literally cost thousands of euro’s just because the switch from Agent 2 to Agent 3 took place.
I understand there are upgrades, and that it’s best to test with your members, but make sure this is thoroughly tested before it’s put on the market.
Perhaps it’s an idea to assemble a team of Replit users to do the testing. And don’t make them pay in this role, let them develop their own platform/app. Do this on a rotating basis.
(I volunteer.)
2 Likes