Hand-coded-only maximalists, explain to me why
"just change that one config line by hand"
is better than asking an agent to
"update the queue size from 3 to 10, and make sure it doesn't cause side-effects or resource contention. keep docs up to date?"
Most of the time, a single file change is probably okay, but why wouldn't I want to have a system to investigate, every single time that change is read:
- how it would impact the system at large and under load,
- where in the documentation that new value needs to be updated,
- what potential race conditions this might cause...
Agents are smart enough to figure all of this out, and they do it in a fraction of the time that we can. Instead of hoping for the best by hand-editing a file, I would very much like to give it to an agent that has full context of the whole codebase and can run every single test we have available.
So often, a small change has had massive cascades that I didn't expect somewhere down the road. This has happened significantly less often with an agentic system that has 100% focus and 100% context every single time it runs.