I wanted to bring up something I’ve noticed while working with TypeScript on Replit.
When I rely on the built-in LSP, it consistently shows 0 issues in my project. However, if I run:
npx tsc --noEmit
I immediately see 500+ errors.
At first, I’ve been ignoring this since TypeScript is ultimately syntactic sugar for JavaScript.
The agent was still vending functional JavaScript, so things would run.
But this ignores a key point:
TypeScript’s value isn’t just in catching syntax slips—it enforces strong type safety, which helps developers (and agents) catch problems earlier and write higher-quality code.
If the LSP surfaced TypeScript compilation errors instead of just showing “0 issues,” the feedback loop would be much tighter. Agents (and humans) would be nudged toward writing type-correct code, which would mean fewer runtime surprises and overall better output.
I spent 5.3 hours working with Agent 3 to resolve over 500 TypeScript compilation errors. Over that time, I spent $71.58 in agent costs, which comes out to about $13.50/hr.
At first glance, that’s not bad — it’s certainly less than what I’d spend on one of my Sr Engineers to chase down that many compilation errors over a sprint or more. And the result is positive:
My repo is now more stable.
Adding features moving forward should be more efficient and less error-prone.
Here are some positives I’ve seen with Agent 3:
The orchestration layer feels well-implemented (not sure if it’s built on PromptFlow, agent-to-agent coordination, or another framework, but the underlying execution flow works smoothly).
I appreciate seeing the Agent explicitly engage with the architect role — that separation of concerns gives more confidence in the reasoning process.
I’d like to see the “Thoughts” logs persist in the chat, so the interaction between the main agent and sub-agents is fully visible and can be reviewed after the fact.
Overall my experience has been fairly positive, though I recognize others may have had different mileage.
I’ve seen reports of Agent 3 usage costing ~$50/hr, but that was not my experience today — my effective rate was much lower.
From my perspective, Agent 3 represents a significant step forward in both accuracy and efficiency compared to earlier iterations.
But here’s the issue:
The only reason I had to burn those hours in the first place is because the Replit LSP doesn’t surface TypeScript compilation errors.
If those errors had been caught up front, the agent would never have generated such a mess, and I wouldn’t be paying out of pocket to clean up something caused by a tooling oversight.
So even though $13.50/hr is cheap compared to hiring an engineer, the principle still stings:
I’m paying to fix a problem that Replit’s product could and should have prevented.
Here’s a constructive idea: instead of refunds (so it is not hurting the company), Replit could offer credits when users encounter situations like this. Not necessarily paying me back with $, but:
Credits cost Replit nothing in cash.
They let me keep momentum on my project.
And they show goodwill to the community, who are frustrated about paying for preventable issues like this.
Even at a relatively low hourly rate, it doesn’t feel fair to pay for work that’s only necessary because of missing features in the platform.
I still don’t see the point of assistant - never have. Get your senior developer (agent) to build your app, then ask assistant (junior intern developer) to fix it. That is nuts to me, but I am clearly missing something about what assistant is truly capable of. I wish someone would tell me.
I respectfully disagree with you. See my comment on your thread, that I made about meta prompting. $.10 - $.50 for the Agent 3 to make a user prompt that I then give to the assistant to implement at $.05 is the the way to go IMO.
I suppose I woke up this morning (in a very wet and windy UK) trying to think how we can find a way that everyone is going to agree on. I simply don’t really know any more, if I am honest. The right model is going to be an utter nightmare to work out.
I think we’re all agreed that they rushed into Agent 3, pricing changes, and the new workspace, and it is all an utter disaster. Lucky they got that new $250m in the bag before the new shareholders saw the feedback hey!
The only thing I can really add is: if they sit in a darkened room without all this constructive input from real users, then they will continue to get it wrong. Hopefully they will engage with us and start to make some headway with a model that will satisfy a majority of users. Roll on Agent 4.
EDIT: Sorry, I just realised I did this reply on the wrong one of your posts. I think we are all ranting in so many posts now, we’re losing track
Well, it still hurts because they’ll have bills to Anthropic, but yes it does help cut back on processing fees, etc.
However, some way to help bring projects to “current standards” would be helpful. This new agent thing certainly reminds me of what the future equivalent of patching and upgrades is going to look like
I love Replit more than others I’ve tried, and trying to stay optimistic, been a rough week this week. Hopefully they come out of this with strong lessons learned, and come out with processes that prevent this kind of product launch again in the future.
Linear cost increases should happen as functionality grows and can support more enterprise-grade use cases. Exponential increases cause people to pump the brakes and lose trust, and am really hoping the message was heard on this.
In terms of product functionality, I’m hopeful that “You can never make the same mistake twice, the second time it’s a choice” applies here and there will be better testing done prior to launch, preferably on old code from the previous agent and a Replit official statement on best practices to handle converting old projects to the new launch.
You sound like a British politician That’s all they ever do, is say “lessons have been learned”, and then they trundle on to make the same mistakes 6 months later.
Let’s hope Replit really do learn their lessons after this.
