Hey everyone. I know Replit’s 10-year anniversary is tomorrow (May 2nd) and they are giving away service for a day. I’m posting this because I need to warn people before they attach a credit card to their accounts to keep using the AI agents after the promo ends.
I want to preface this by saying I have loved Replit up until recently. I'm a relatively heavy user and normally hovered around $1k–$2k a month in spend. I even used to buy the $1,000 credit packs in advance to save money because I relied on the product so much. Before the recent Agent 4 update, the AI was fantastic and well worth the money. You asked for a feature, it did the work, it stopped, and you moved on.
But the new version introduced this "plan-while-building" architecture where it spawns multiple parallel sub-agents on a Kanban-style board. When it first started happening to my projects, I was caught completely off guard. I approved the first few dozen tasks assuming it was just mapping out the steps to complete the small, scoped features I asked for. I didn't realize that approving one task gives the system permission to endlessly invent new ones.
I got caught in a sunk-cost loop. I was terrified that stopping it mid-stream would wreck my codebases and leave my architecture broken, so I thought if I just let it ride, it would hit 100% and stop.
It never stops. I literally tracked the stats on one of my projects: I asked for 23 things, and the system generated 770 tasks to deliver them. On another project, the daily amplification ratio was 11.3x. For every 1 task the AI actually merged, the platform auto-generated 11 brand new follow-up proposals.
I'm posting this constructively because I know Replit staff read this sub, and as a paying customer I need them to know there are some massive architectural flaws right now that are actively billing users for platform bugs:
- Forced follow-ups: When an agent finishes a job, a built-in platform skill forces it to propose 1-5 new tasks. We can't opt out of this.
- Billing for parallel collisions: Two agents running at the same time don't coordinate. They will propose the exact same follow-up task, and Replit charges you twice to build it. If they touch the same file, they cause a rebase conflict, and the platform forces the second agent to re-run all its code gen and testing just to fix a conflict that the platform's own merge ordering caused. And we pay for that compute. Replit's docs say this isnt the case, but I can promise it does indeed to tons of unnecessary work.
- Paying to fix its own E2E pollution: The AI runs E2E tests, pollutes shipped data files with dummy data, merges it, and then the next agent spends paid time (and our money) reverting the pollution the previous agent caused.
I am completely stuck. My Git history is a minefield of hundreds of automated commits. I don't even know what commit I should revert back to because my legitimate feature requests are buried inside a $6,376 avalanche of auto-generated tasks (way beyond my usual monthly spend). My codebase is basically being held hostage by a system I can no longer afford to run.
The worst part is you can't even stop the queue. There is no bulk "Cancel All" button. The API only lets an agent cancel its own follow-ups, so once an agent exits, those tasks are stuck in the queue forever unless you sit there and manually click "Cancel" 130+ times while the active ones keep spawning more.
I've submitted detailed support tickets with commit hashes, task IDs, and logs proving this is a runaway platform loop and not user error, but I haven't heard back in days while my bill just sits there at over six grand.
If you are logging into Replit on May 2nd to take advantage of the free day, enjoy the sandbox. But the moment you attach a credit card for a real project after on May 3rd...:
- Do not trust the sub-agents.
- Never run them in parallel.
- Scrub your task queue constantly.
- Set hard budget limits immediately because the platform will not pause itself.
I'm really hoping someone from the Replit team sees this and can look into my ticket, because right now the Agent 4 loop is a runaway train. Be careful out there guys. I know there will be folks that are quick to criticize (its the Internet after-all) but mostly wanted to post this as a warning to others.