My feedback after using an agent everyday at work for two months
Two months of daily coding-agent use: faster delivery, shifted cognition, new strengths, and an uneasy sense of dependence.
It’s been two months since I started using Codex at work, everyday, to “assist” me in my development assignments. It’s time to write down some observations. I like to refer to the agent as “Rocky” like in the “Project Hail Mary” book/film.
Observations
- I’m lazier when it comes to manually understand the intrinsic aspects of the codebase, how certain things work etc, I want to (and in the end, after a while I do) but I rely on Rocky to explain them to me, instead of grinding it myself, so I guess I’m deporting some cognition to Rocky, which is problematic.
- Overall I’m better at seeing the bigger picture in terms of architecture. Not sure if it makes up for the 1st point but it’s something.
- At the end of the day I’m just as tired as I was when I coded manually, nothing changed on that aspect.
- I deliver approximately 2 to 3 times faster, which feels good in general, like being on a really fast straight line rollercoaster. But it feels like cheating.
- I feel a bit weaker because I wouldn’t do my job at the same pace without Rocky. I need Rocky… and with need comes the loss of strength.
- If the problem is longer to explain than to fix, I fix it myself. But when I do it feels really slow. Like when you drive the highway at 120 km/h and then you slow down to 80 km/h, it doesn’t feel like 80, it feels like 40. Well it’s the same when I code myself. Altough it feels nice to tackle problems like in the old days.
- It’s a rollercoaster of feelings, sometimes I’m highly confident and satisfied, and sometimes faillible and doubtful. That makes me think of the mechanism of addiction. I guess.
Conclusion
The uncool part is that the obligation of using an agent to work is implicit: the roadmap is so dense it’s inevitable. But maybe the worst problem we have at work is that we still have deadlines, duh.
The worst part of using Rocky is the offloading of cognitive workload and manual tedious exploration to an external source. The whole challenge I think is to keep the productivity gain without renting brain effort. More easily said than done. Or just stop using an agent like that, plain and simple. Or accepting that this particular cognitive effort I have trouble letting go is from the past and focusing on finding it elsewhere, like harnessing the agent or whatever. Troubling times my friends, troubling times indeed.