Beyond the initial build, Agent is truly a mess. You’re so much better off forgetting that Agent even exists after the first prompt is complete and just using Assistant.
Am I alone here?
6 Likes
you are right - assistant is much better.
saying that, i might try agent if assistant is getting nowhere
2 Likes
Oh mehn. You’re right. Agent is just dumber in the long run.
It’s basically useless after making the initial database
I think agent can throw together some nice UI stuff quite quickly, but I often find it’s left me with a directory full of files that I can’t quite follow too clearly. And man, it’s 5 times more expensive that assistant, any half the time you end up spending that on just asking it to retry or correct things.
1 Like
You down like 2 dollars in credit on a simple error. Sometimes i feel it does it on purpose
2 Likes
Course it does yeah. When you use the assistant as well, try giving it like 20 things to do to try and get your money’s worth out of the 5 cents, and watch how it mysteriously forgets to implement some of them towards the end of your list. There’s definitely some limit on that thing. But these guys have got to make a buck I suppose, so fair enough.
Yh business as usual. But what i would like. Is for the agent to communicate in simple English the code changes it has made IN DETAIL after each prompt.
Assistant explains well to an extent but still not good enough
Is Agent running on Claude too?
I have started to use Agent to build the bones, then using Claude alongside afterwards(not assistant as I find it janky)
yes but didnt used to be this way. Agent used to be great for larger tasks ie adding complex additions to a site or codebase, but recently I feel the same as you do. Agent just tries to start my app build all the way over while ive actually had very little issues with assistant, even on traditionally complex issues.
You use Agent to build the “House” structure. You use assistant to do the detail work inside the house
1 Like
I honestly think mostly because we use assistant mainly for the smaller tasks so it’s just not about if one is better than the other, it’s probably just dependent on what tasks we are trying to accomplish.
1 Like
I feel “Agent” is like a heavy hammer getting the big initial things done; and “Assistant” is like the fine tuning hammer coming in to refine and provide detail. I spend most of my time in “Assistant” but go back to “Agent” when “Assistant” is not able to get things done.
4 Likes
i agree with everything you wrote
If I can I just use Agent for everything. Yeah I get errors but I just copy and paste the error into the chat without any comment and it usually fixes it.
1 Like
Sadly, your time will come.
sometimes I burn like 5-7 checkpoints on one error but I don’t really care about it. it ends up being more cost effective and time saving than using assistant, in my experience. my last project that I built using only assistant it took 7 iterations and the 7th iteration took 355 checkpoints to go from zero to working. I feel like agent could have probably done this in 20-50 checkpoints. the big difference there would be time, this project took about 20 hours, I don’t care so much about the $5 cost difference in 50 agent checkpoints vs 355 assistant checkpoints but if I could have done the project in 5-10 hours compared to 20 that would be cool.
1 Like
Yo @whereisthestart, do you have a clean way of using other LLMs on your code? I found Xai is good when I get stuck but I have to switch platforms and explain all the context every time. Ends up saving time getting caught in error loops but it would be cool to have a cleaner UI. I like getting different LLMs to bounce ideas off each other and gain consensus on the best way forward. Like an LLM brainstorm.
but if it makes a checkpoint, it charges more vs the Assistant
that’s technically true although i find that agent is often capable of doing more in 1 checkpoint than assistant is capable of doing in 5 checkpoints, so it ends up being more cost effective to use agent until you get to the point where you’re making requests like “decrease the padding by half” or “change the blue button’s color to red”
Not really, I set up a project in Claude and copy in code / errors as I hit them. Helps stop Agent getting stuck in a loop
1 Like