I was talking to Kody about this – to me it’s really far too complex and there should be just one. You don’t see other chatbot systtem with two “kinds of” chatbots, and the choice “which do I use?” is a total UX nightmare at the moment. Basically this is what I’m doing now (again thanks to Kody’s patient 1:1 chat!)
I use the Agent to start things off. I ignore all the “add these features?” options. They’re a siren song to misery as the scope is larger and it will probably fail.
once it’s going I switch to the Assistant and stay there.
Before anyone from the team responds to this: do have a pause to appreciate the mental model you have (which is insanely complex and integrated incrementally over months) vs a new user (who just wants to say stuff)
I understand it’s hard to scope it but it’s also not the user’s concern that this is a problem – the PM on this product should be obsessing about this relentlessly.
FOR EXAMPLE – honestly, if the Agent did the first screen “what do you want to make?” and just switched permanently to the Assistant inside the IDE people would probably have a lot more luck.
I think I understand why Replit had to go with it. Agent is more capable, especially when it comes to the initial setup which involves a lot of steps and can really impact the success of a project. A more capable agent costs more to run, so it will likely have to stay separate. It’s like using GPT-4o vs GPT-o1. One is great for most tasks and cheaper, the other one uses advanced reasoning and is more expensive.
I totally understand your point of view and I do agree with you that after Agent builds the initial prototype, Replit should automatically switch to Assistant or at least remind the user about it. One time I forgot and I kept asking Agent trivial questions, but paying the higher cost for each prompt.
Here’s my question about Replit Assistant and web scraping:
I noticed in the description that both Agent and Assistant support “website scraping,” but when I use Assistant, I don’t see web scraping capabilities like I do with Agent. Can Assistant actually perform web scraping? There seems to be a discrepancy between the description and my experience with Assistant.
Correct me if I’m wrong, but Assistant only makes one-time changes across all files, whereas Agent is capable of autonomous iterative development. I actually use Agent for most feature development, whereas I use Assistant for bug fixes or style changes or changes to the environment (i.e. user auth persistance).
I agree with the previous post about knowing when to reach for agent vs assistant. The mental model that I have been using is to use agent for anything related to new fields or features that require API changes and updates to the database. I use Assistant almost exclusively for UI changes where I’m changing the screen but not impacting the data.
I’ve found that Assistant will usually miss required imports and will not look through a whole file of missing imports when asking to fix the code. Agent doesn’t seem to make these same mistakes as it gives a more comprehensive evaluation and fixes everything at once.
Long term I agree this should become a single agent with some sort of orchestration at the point of request to determine who to task to do the work. No different than a PM who gets a request and determines who they should task to make the change. Junior or Senior? Front End, Back End or both? Much easier for the end user and helps Replit with cost containment.
That’s fair. To me it’s a matter of tradeoffs. If money is a major limitation it would make sense to figure out how to maximize the use of both Agent and Assistant given the price difference. If the assistant gets it right in 2 tries, then it’s 40% cheaper. If money is less of an issue, then agent for everything makes sense as it’s faster and typically more comprehensive.
For those that have zero coding experience, and cost is a concern, they would want to reach for the Assistant for a prompt like “move field x above field y” or “make the background of this thing light gray instead of white”.
I observed an interesting phenomenon during an experiment: I built a simple app with registration, a database, and some interaction flow. The agent performed flawlessly until about 80% completion, then began looping on the same issue. I tried different prompts and approaches without success, so I switched to the assistant. The assistant immediately understood the problem and fixed it on the spot. However, when returning to the agent, it seemed like its memory was wiped, and for the next task, it could completely remap the entire app. This often required rolling back to an earlier state.
I also tried forking the project to refresh and reset the agent’s memory, but it didn’t help. Is anyone else experiencing similar challenges? I’d love to see real case studies and successful prompts used from start to finish for building apps.
I would recommend reading this other mega-post going into a lot of details on the differences between Agent and Assistant and how they are intended to be used: Pair Programming with Replit's AIs
Agree. You need to KNOW the difference between the two versus INTUITIVELY going from one to the other, right?
That should be enough info right there to tell you it needs a rethink.
I think it’s important that from the outset people know the difference in pricing. $0.25 vs $0.05 is a big difference over time so people should use assistant where possible and agent where assistant is struggling.
Eh…I think you’ll find that beyond bug squishing, it’ll take assistant 5+ tries to implement anything new. It doesn’t do a good job of looking beyond the immediate context scope (which is preferred sometimes, for sure) but that means it’ll just simply break things, and you’ll have to copy/paste errors 5, 6, 7 times before it finally works through everything it broke.
It’s probably a wash, and time is money. I find myself just using agent for most things that isn’t strictly bug work or codebase analysis/chat.
Yeah I saw this a bit although one thing I noticed is that sometimes it does actually make the change but you have to either wait or manually refresh the page for it to take effect. Point taken though although the agent is not perfect by any means. Do you find assistant better for more technical issues like deployment, proper errors with pages not loading etc or not really? I do think the difference is confusing.
When I have a persistent issue that agent can’t fix, or it starts to loop, I usually turn to browser console, copy paste in assistant, and it’ll start to unfudge itself. I use Assistant for basic stuff too, style changes, spacing, CSS work, etc. - but I honestly tend to just take it up the tailpipe at 25 cents a wack and work with agent 80% of the time.
