I hit a new personal best for longest running agent session with Agent 4!
Is there anyone else out there who can get Replit to work like this?
If so, I’d love to grab a virtual coffee and compare notes.
Not on replit, not as long, but I had codex do a long running task that was over an hour. The prompt was maybe 1k lines of markdown and the Plan mode took 10min of back and forth confirming things. I did six iterations of the prompt in order to finally get it cooking how I wanted.
*more impressive - that was only $60ish on Replit
I would never let my agent spin for 8 hours straight.
That’s just me though.
How are you able to determine if those 8 hours were spent doing something productive? Not saying it wasn’t… genuinely asking…
Do you manually audit the work? Testing loops?
Letting several parallel agents spin that long without closely monitoring what the heck is going on that whole time is super risk IMHO.
I mean, it’s super sweet that 8 hours only cost you $70 bucks that’s like paying base price for a pair of Nike shoes right from the sweat shop…
I’m just saying I’d love to hear more about your process in written form.
I’m also trying to get a group of Replit devs together for a USAA style Replit group where we cover each others backs. Like when the soldiers who first started owning cars… No one would protect them if they broke down and needed repairs.
They took it upon themselves to make a group, formalize, and now it’s the company we know today. Not saying we’d be the next USAA lol
We could get a group of us going to do the digital version of this. The vibe code era seems very similar to the mass production of automobiles…
Let me know if you’re interested!
I’ll make an application (using Replit) for us to network and connect, start coordinating and getting things rolling if anyone’s seriously interested.
LMK!
Depends on the project. If it’s a production app probably not. If it’s experimenting and trying to get better with AI tool probably. I don’t know if I have any tasks that would take 8 hours. What actually got done @mikeamancuso?
1 Like
Yeah, I’m all for experimenting and learning with low risk builds. To me knowing what I know about the agent, and using AI to build in general .
I would never spend this $ on a session I didn’t know was producing the work of at least a mid tier developer who’s sick, or just cooked atm lol
A low price tag of 7.4 / hour is cool.
If you’re doing that for production apps at scale though I’d really need to know the way you approach the slop it will inevitably produce.
Doing that at scale, to me screams, “high technical debt”.
I personally would not be okay with doing this unless it was an intentional experiment.
@mikeamancuso would love to hear more
how did you get this data generated? @mikeamancuso
Hey everyone, thanks for the thoughtful replies!
That 8h 26m session turned out to be a spectacular success. I had one focused agent stay locked in the entire time on a single goal: running a continuous clean-code improvement loop.
Here’s how it worked:
- The agent pushed changes to GitHub → triggered SonarCloud and CodeQL analysis (smells, warnings, hotspots, bugs, etc.)
- It waited for the results (natural idle time during static analysis)
- Then it read the feedback and performed targeted refactors
- Then it looped back and repeated until the analysis came back essentially clean (“all perfect”)
I do a daily clean-code ritual because I was tired of the 70/30 split: spending ~70% of my time cleaning up after Replit’s “get it done fast” output and only 30% on actual enhancements and new features. Replit’s agent is great for speed and prototyping, but the accumulated quality issues made it hard to push apps past a certain complexity threshold without things getting messy.
This long, intentional quality-focused run completely flipped that ratio for me. No slop piling up, just steady, measurable improvement toward a tidy, high-quality codebase that’s always ready.
I totally get the hesitation around long-running sessions and the risk of technical debt—especially for production apps at scale. That’s exactly why I built the process around SonarCloud feedback and iterative refactoring instead of letting it run wild. It’s an experiment that’s now becoming part of my regular workflow.
Happy to share more details on the exact prompt structure or pipeline if anyone’s interested. Also down for that Replit dev group idea — sounds like a great way to swap these kinds of approaches.
What about you @pmz, @RocketMan, or anyone else — how do you usually handle quality control on longer agent runs?
1 Like