I think AI can reverse engineer a stack easily but software is more than just a fancy UI and even though the benchmarks are improving, I mostly see marginal improvements in accuracy for most of the internal tasks I do.
As long as AI companies are not willing to accept the outcomes of worst case scenarios, I don't think models can reverse engineer any company/outcome.
Models are still very far from solving the absolute corner case scenarios like privacy, compliance, availability, consistency, reliability, disaster recovery, SLAs and so much more and this is where software engineering gets super interesting.
The other day, I was looking at a scenario where a spike in agentic traffic led to lots of 5xx. I iterated couple of hours with Opus and the conclusion ? Add more nodes (which is Indian equivalent to restart and see if it works). Also why would I spend $100 tokens to get to this answer.
I’m increasingly convinced that coding agents will become less like copilots and more like junior engineering teams.
You give them a goal.
They break it down.
They implement.
They test.
They review each other’s work.
And eventually, they open the PR.
I think the next evolution of coding agents is moving from:
“Give me a task and I’ll write the code”
to:
“Give me a goal, figure out the subtasks, run multiple agents, test everything, and give me a PR.”
The interesting problems are mostly around orchestration.