I keep hearing about how coding agents have gotten so smart and everyone is posting their 12 agent setups and what not, but for some reason I feel like I haven't found the one habit that actually changed how I work. I still use Claude and Codex daily but still.
So curious to the people here, what is one thing you started doing with agents that changed your work for real?
At this point, how much of looking productive at work comes down to luck?
Good model that week, good task for an agent, a clean part of the codebase, a reviewer who just approves. It often seems to depend as much on what you got assigned as on skill.
I feel 70-80% of devs can use agents reasonably well now. Most people can prompt, plan and fix what comes back when they get a task the agent is good at.
Is the game becoming more about getting the right tickets at the right time?
Every model goes the same way. Week 1 it cooked, week 3 it got nerfed, week 5 a new model drops and the cycle starts again.
This time I saved 10 PRs from this week, prompt and diff, so when the nerf posts start I can run the same prompts again and actually check instead of going on vibes.
My bet is 12 days. Whats yours
We were a codex team since last year. On Tuesday two people tried 5.5, by Thursday everyone had switched and nobody asked me first.
What I noticed from the PR side after 3 days:
- PRs are smaller. Average went from around 400 lines to around 250
- it stopped adding helpers and abstractions nobody asked for
- the PR descriptions actually match the diff
- we burned about half the usage we did on Sol for the same amount of tickets
Only downside so far is it asks more questions before starting, which some people love and some find annoying.
Did anyone else switch this week? And is anyone staying on codex, and why
For months my days looked like this:
me: add retry logic to the upload endpoint
opus 5: Done! All changes applied and ready for review.
me: the endpoint file isnt even in the diff
opus 5: You're absolutely right, that's my mistake entirely.
Three days on 5.5 and that has happened zero times. It still misses things, but now it ends with a list of what it didnt do and why, which is all I ever wanted.
Anyone else seeing this or am I still in the honeymoon week
I wrote a small script by hand on Saturday, no agent, about 80 lines. Took me 2 hours and it was the most fun I had with code in months.
At work my day is give the agent a ticket, wait, read the coderabbit summary, check the flagged lines, approve. It works and I ship more than ever. I just dont really make anything anymore, I supervise.
Do you still write some of it yourself on purpose? Or did you make peace with the supervisor job
We switched on Tuesday like everyone else. The speed is real, a ticket that took 20 minutes now takes maybe 8.
What I cant tell is if the PRs are better or just arrive sooner. Coderabbit was leaving about 4 comments per PR last week and about 3 this week, which could be the model or could be that everyone picked easy tickets to try it on.
Is anyone actually measuring this? Or are we all going on how it feels in the first week
My whole setup is: open the terminal, prompt it, read what coderabbit says on the PR, merge. No subagents, no skills folder, nothing running overnight.
Meanwhile half my feed is people with 14 markdown files running a fleet of agents. I tried the long running thing once on a refactor. I came back 3 hours later to a 60 file PR and spent longer checking it than it would have taken to do it in small pieces.
Am I underusing this or is most of that setup for show? What does your actual day look like
[ Removed by Reddit on account of violating the content policy. ]
[ Removed by Reddit on account of violating the content policy. ]
Trying to get a real number before I go to finance. Right now we are at about 200 a dev for the agent plan, 24 for coderabbit, and some API overage nobody tracks. Call it 250. The agent side is where it moves, one guy burned through his whole month in 9 days.
What are you at per dev, and is anyone actually tracking the overage or do you also find out on the invoice
12 years in and I still write most of my code by hand. Not against agents, I just have not found a way to use them where I understand the PR at the end. Every time I try, I get 40 files, a description I dont trust, and an hour of reading.
Meanwhile people on my team ship 6 agent PRs a day. I asked two of them how. One does not check the output at all. The other has a whole thing with a plan step, then the agent, then coderabbit, then he only reads what got flagged. That one seems sane but I have not made it work for me yet.
If you got real quality out of this without reading everything, what is the process? Step by step, not vibes