i wrote this while claude wrote the code.
its 3am. fix agent is running for the third time, i have been waiting on it more than an hour now. no progress bar no ETA nothing, just the spinner. and i am sitting here with nothing to do because there is literally nothing for me to do till it finishes.
i have been reading this community for long time, silent reader, never posted. this is the first one. i think people here will understand this better than anyone around me right now because everyone else only sees the output. "you shipped this in one week??" yes. we did. but i am not sure anymore who is "we".
remember the whole thing. robots will do the hard work, humans will do the art.
ok. robot is doing the hard work. and i guess this is my art, a blog post at 3am while the machine writes the thing i used to write.
the project - its an AI control tower for a logistics client. it keeps drivers and vehicle owners notified proactively, tracks the trips, and one of the main features is it detects when driver is getting drowsy and alerts him so he doesnt sleep on the wheel. accidents prevention basically. good product, i actually like it, thats not the issue. and the code is good also. no doubt about it, honestly better than what i would write at this hour.
but its not my code. thats the issue.
llm was supposed to help devs. and it does, i am not going to lie about that. but what happened is the bar got raised. now the expectation is one week for what used to be 2 months. and whatever work is left for me is the most boring part of the whole thing and i am doing that for even longer hours than before.
boring meaning, no thrill. no problem to solve. only thrill left is solving the business problem, and ya that was always the point i know that. but earlier we had small wins on the way. big bugs that keep you awake because you want to be awake. fixing the unfixable. that feeling when you finally see it and go ohhhh and everyone in the room knows. thats mostly gone now and i think we will never have it again like before.
when was the last time you even read the error properly. like actually read the whole trace and went and fixed it yourself. for me its been a while. there is no time for that, we are operating on top of code where we have not read or understood 90% of it. we are not devs on these projects, we are like supervisors of something we dont fully understand.
and then the 5 hr limit.
for complex stuff you basically have to plan your sleep around when the limit resets. run a phase, hit the wall, wait, run next phase. i ended up purchasing another premium with my other id just to keep working. then also i took a friend's account for a day. i think a lot of people are handling 2-3 subscriptions right now quietly and its exactly like managing multiple resources on a team, except the resources go to sleep on a fixed schedule and you dont.
and the agent stalls. one hour into the run. no error no timeout, just stopped. and you are sitting there saying just get it done bro. to a text box. at 3am.
never felt this useless in the process of coding before. not as a fresher, not on my worst day. useless is the word. the thing got built, mostly by claude, got deployed, launched to the customer, real drivers are getting alerts from it. and i am living in fear waiting for it to break because i know when it breaks my only option is to go and beg claude again.
ok i dont want this to be only a rant (may be it is). some things that actually helped in this week, if you are in the same place -
make phase wise plans before you write a single prompt. not "build the app". phase 1 auth, phase 2 data model, phase 3 apis like that. each phase small enough to finish in one session. this alone saved a lot of the 3am pain.
write handover notes after every phase. what is done, what is half done, what is broken, what next phase needs to know. i keep it in a plain md file in the repo. when context resets or limit hits or you switch account you just paste the notes and continue from where you left instead of explaining everything again.
read the error yourself first. even if you are going to give it to the agent anyway. 2 minutes, actually read it. half the time you will see it. other half at least you know what to ask instead of "fix it" for the third time.
debugging is the real skill now. not typing code. if you cant debug then the first time something breaks in prod you are fully at the mercy of the model. protect this skill, use it on purpose even when you dont need to.
know your reset time. sounds stupid but plan the heavy phases right after reset and do the reading and planning while you are locked out.
keep at least one part of the codebase that you understand end to end. for me its the deployment and the db layer. if everything is on fire atleast i have somewhere to stand.
so is this it? i honestly dont know. client is happy. thing works. by every outside measure this was a big win and i should be feeling great.
but i keep coming back to - what was my contribution. better prompts? staying awake? knowing which account still has hours left?
may be this is just what the job is now and i have to make peace with it. may be the thrill comes back in some other form which i cant see right now. may be in 2 years this post will read like someone crying about compilers.
or may be lot of us are feeling this and not saying it because the output looks so good that complaining feels ungrateful.
anyone here also did the same "one week" thing recently? what's your view?
agent finished btw. 3:40. going to sleep.
Top comments (0)