I was reading Amjad’s thread on Twitter and it felt like those of us complaining about the price hikes were a minority. A lot of people on there seemed happy with the Agent 3 update (albeit there were still those upset with changes).
Reading the thread, I felt coming out of it they are making the statement “Previous Agents wrote a lot of bad code, Agent 3 is so good at what it does it wants to bring everything up-to-date. The time it invests is costly because it has to do so much to bring your code back to best practices.”
I hit the brakes since the day of the release until yesterday to calm the emotions. Now I’m trying to think of this objectively if this is, in fact, true. Has anyone just ran with it, let Agent 3 do it’s thing, and overhaul what they’ve built?
I’d get if agent 3 was struggling with agent 2 apps. But if agent 3 is no good at starting new projects from scratch either, then Replit have a real problem. All I can say is bare with, I am positive they are reading all this and head down in their computer screens fixing stuff.
I let Agent 3 run for about three hours to resolve 500+ TypeScript errors, which cost roughly $70.
I’m not sure whether it also addressed any technical debt during that time. Given the reports of Agent 3 running at around $50 per hour, I have to assume it didn’t.
Overall, I was satisfied with the work it completed. I do have a technical background and stay very hands-on, so I was able to follow and assess the process closely.
I am confused. You say you let it run for 3 hours. But also said you were hands-on - how? Once agent starts running we let go of the steering wheel and have no input.
Sorry, but there is no way on God’s green earth I am letting an agent run for hours at a time. 10 mins is about all I feel comfortable with.
Someone else on here used this analogy, and it’s a good one:
Like getting in a taxi in a new city, and the driver refuses to tell you how far away your destination is, how long it will take, or what the billing model is (is it per mile, or measured by time, or both?).
I had it iterate on the output of npx tsc --noEmit until there were 0 errors and monitored it.
If you look at my LSP post, there are several screen shots of work/cost. Each one i had to interrupt/course correct it, until the final screenshot where it understood what I wanted and I just let it run.
Was this running with Max Autonomy? How many lines of code did this end up being? Link to your post here if possible.
I’ve started running it cautiously. My prompts have been running 20-60 minutes with costs around $5-7 (I’m much happier with that cost vs. double that). Similar use, feeding it errors for it to fix, mostly around PUT/POST requests.
I did link it. Look a few replies up. The screenshots in that linked post have all the info. It’s not the OP, but scroll down to my second post on LSP thread.
