I also learned you’re supposed to rollback any changes if they don’t work.
However, Sometimes it just can’t figure it out…. And neither can I. 
1 Like
Is Assistant able to work more broader use-case? I think I found the Assist feature falls when you go for covering a use-case level impl, like secure all my API’s? Anyway I have equal frustration with both uncovering deep bugs… i still have few that no one can solve 
Replit highly recommends using the Rollback feature rather than having Agent/Assistant undo changes. They’re very insistent on that. Like 11 on a 1-10.
So, when Assistant/Agent ask for your approval for code changes, you can go back post-changes and the button will now say something like ‘rollback to this state’ or something like that. You’ll see.
Hope that helps!
1 Like
I have only used Agent to iterate over the prototype and was able to gradually build features. Agent did a good job with building decent frontend and able to refactor to add more flows. I did leave some unused files.
A guy on my team can help you figure out the tough stuff.
I can’t figure why this is the case; Cursor seems to have claude on its agent work and does such a better job. Replit agent goes around in circles. Really, if it fails, it should not be charging is as users. I’ve wasted so much credits on this issue.
1 Like
I could be wrong, but I think it’s because of how they’ve programmed Agent for token stop gaps and billing. I honestly think it spins it out instead of maintaining a solid workflow. It constantly interrupts itself to encourage new chats (to reduce their contextual costs). It continually makes unnecessary checkpoints to restate what you said and ask if it should continue (which you already told it to do). All of this leads to lost context and, ultimately, more bugs and BS.
They have a tough challenge: maintain profitability while letting the AI go nuts and do what it needs to do. I don’t think many people here understand how expensive it is to send max-token API requests to flagship LLMs. So there’s a disconnect between how you all feel (and often, me) and the reality of the costs of running this type of software.
2 Likes
Yo @realfunnyeric, you have a guy that can help with the tough stuff?
I just finished posting a new topic to the effect of ‘when do you restart’ or something like that. Feel free to check that out for details of where I’m at.
When do you scrap a project and restart? - Agent & Assistant - Replit Community Forum
1 Like
That is an interesting observation. I built my first app (www.dashdiary.co.uk) using the agent at the end of December 2024. Whilst there were a few kinks to iron out, the process and results were good. I have subsequently been working on two projects and the agent just cannot seem to get things right, to the point that it feels like it is purposefully wasting money.
On one project, it cannot get the Firebase authentication (or authentication state) working. Something that was a breeze in the first project.
On the other project, it simply cannot correctly apply a search or dropdown filter on a Postgres data set. Again, something that was a breeze on my first project.
I wonder if Replit has changed something since December 24/January 25 that is causing these issues?
Even the initial build!! It can’t even properly booth my backend for my APP in Replit environnement… Hours of running in circles… It’s a fraud!!
Definitely not a fraud.
I’ve launched multiple apps, some of which are pretty impressive.
But there is a learning curve. And to say it’s for non-techies is disingenuous. You absolutely need to understand dev principles, at least basics, in order to get yourself out of the numerous binds that happen.
And knowledge of Git, and how to properly use it, is mandatory. And it’s fairly complicated. Take the time to learn it.
1 Like
The main use beyond the initial build is that its fun for kids to play away with the Agent. Primarily focused on the Assistant myself too… in the past we were able to choose various models to use for the assistant too, hopefully the bring that back as I’d like to see how xAI / Gemini perform vs. Anthropic… not sure if its the best for code generation anymore…
CEO of Replit told me they’re just waiting on API approval for Grok 3.
1 Like
Eh…not really true. It’s great at adding new features. As with all things AI up to and including this product, it’s about your prompting. It has it’s drawbacks, but if you learn how to prompt it correctly, and understand baseline dev framework, it’s still fantastic and revolutionary.
yeah - I see your perspective here, for me personally I find assistant easier to keep on track with the objective(s) I have in mind, that is the step by step approval process forces me to breakdown the process (which is likely based on my prompting). The agent though is great when early on as it can handle the back and forth to get a concept completed (user auth, etc.). The assistant mode also keeps in more and helps me with knowledge gaps (e.g. creating tests).
That’s great to hear about waiting on approval for Grok 3 API inclusion; another feature I really like is the ability now to select/deselect assistant recommended actions and the fact it reversed the flow of updating database updates after making schema changes, reduces friction even further.
There shoud be a “middleware/hub agent”, that decides what to pass to Agent and what to pass to Assistant. The user should not have to make that decision.
Absolutely agree. I had too many instances of Agent deleting and changing working code that was outside the scope of request. It’s like a fox in a hen house. Very unpredictable. Have 3 projects in heavy refactoring mode due to Agent related issues.
Doing this classification is actually an extremely complex technical challenge, that’s why everyone who does AI code assist/agent has the User do it, not just us. You get an enormous amount of false positives where the model would try to select the agent for assistant-y things which charges the user 5x as much and the user gets upset. User selection of an opt into the higher priced tool is important. Could be some product improvements around it though.
Agent takes actions across multiple files / contexts, assistant scopes down to ind. files and acts there. Once your project gets really big or becomes interrelated it’s harder for agent to touch one thing without breaking something else because we have to choose and limit what goes into its context at different steps.
There’s some improvements to this coming soon especially around letting the agent trace through when it makes a change in one file ensuring it updates in other related files without entering into a circular loop. Should make the whole experience much much better.
2 Likes
This sounds GREAT…and necessary. Breadcrumbs.