I'm sorry to say this but... Agent 4 sucks

Agent & Assistant
1

I actually wrote a long rant yesterday but being a genuine Replit fanboy decided to hold my tongue until I gave it more time today but it just really does suck.

I have never had such a bad experience with Replit as I’ve had trying to do the absolute simplest of tweaks on a finished (and very simple) app.

The latest? Adding a toggle switch but realising it wasn’t saving the changed state to the DB. Why? And I quote: “The fields are completely missing from the backend routes — they’re never being saved to the DB. Let me look at the update route:”

It hadn’t even created the paths to DB!

You guys really have to do better.

TFS

2

Yeah, I’m at a standstill until they fix or rollback. I sucks right now. This happened between Agent 2 and 3, took them about two weeks to get all the bugs worked out before the agent started feeling like normal again. Ugh.

3

By the time they fix the bugs, Agent 5 will be around the corner.

This is my biggest issue with Replit right now, but it’s not really a Replit problem. It’s an AI moves incredibly quickly problem.

Replit’s problem is that they haven’t hired enough humans to make these transitions smooth.

It’ll take several more rounds of private capital injections and multiple more versions/LLMs (like Claude) to produce more robust models.

Each time a new round of LLM model upgrades happens, Replit will follow suit. However, they’ll have to figure out how to integrate multiple LLMs into their agent.

This naturally creates a huge gap in the ability to produce “ready agent models.” The humans at Replit can’t keep up with the humans from several LLMs and much more capital injections.

They have to keep pace with them. They need to find a way to integrate these model updates on the fly, which requires R&D predating AI’s public release (like GPT).

They’ll continue to release agent models with about a month of kinks to work out. Then, by the time the agent model is perfected, they’ll have Agent 5.

Until they invest in having the human staff that can keep up with the speed at which LLM models are upgrading, we’ll continue to experience these pain points and bottlenecks.

Agent 4 truly opened doors with parallelism, and that, along with the magnitude of the LLM updates (all around models just had a huge upgrade, to be honest), is what’s created more bugs than anyone would want.

I agree with Jeff. It’ll be something you’ll have to get accustomed to. When they release and force new models, I recommend planning your time for about two weeks to a month of R&D. Let them figure it out.

When the agent is done being worked out and you have a small gap between Agent 4 and 5, bust out your work. Now faster, and repeat the cycle.

Hopefully, they’ll realize that this gap is causing about a month of turnover rates, essentially, and they’ll address this at some point.

Probably not until they get 15, lol.

It’s like watching the iPhone be released.

When they start making the equivalent of iPhone 16 Max Pro versions of the agent, we’ll be lit, fam.

4

Biggest issue is they need to keep the previous version of the model live until the new one gets its wings, they have the feedback here, on Reddit and I’m sure in their support systems. It’s really horribly managed rollouts and it has be getting me stopping work for weeks at a time while the sort it out.