Back in 1980s there was a time when “real programmers” looked down on compilers. Assembly was more efficient. Tighter. Smaller. Compiled code was bloated, slow, and wasteful. Serious engineers were supposed to control the machine directly.
They were not entirely wrong, but that was never the point.
The point was that computers were getting cheaper, software was getting bigger, and human time was becoming the expensive part. FORTRAN mattered because it proved compilers could generate code good enough to beat hand-written assembly on overall economics, not on every routine. Hardware evolved with that reality, and the industry stopped building software as if humans had to talk to the CPU directly.
Now we are watching the same movie again. Agentic coding is getting the same criticism compilers got decades ago.
- The code is worse
- It makes dumb mistakes
- It is slower once you count fixing it
- People using it won’t understand their own systems
- Real engineers should write code themselves (my favourite one)
Again: not entirely wrong. And again: beside the point. The people making the anti-agent argument sound a lot like the people who once insisted that serious software should still be written in assembly.
The question is not whether an agent writes code better than the best human engineer on their best day. The question is whether manual coding remains the default way to build software. I do not think it will.
Yes, today’s agents are still rough. A majority of developers do not yet use AI agents at work regularly. Not yet. But the direction is obvious.
Anthropic’s analysis of 500,000 coding interactions found that specialist coding agents are used much more for direct automation than normal chatbot-style coding help. In Claude Code, 79% of conversations were classified as automation rather than augmentation. That matters. It suggests the usage pattern is already shifting from “help me code” to “go do the work.”
That is the part many engineers still underestimate. Compilers did not win because they produced perfect code. They won because they moved the human up a level. Agents are doing the same.
Human role moves upward from “write every line” to something closer to:
- define intent
- set constraints
- review output
- run tests (after AI tests itself)
- fix parts sometimes (still needed, same as Assembly, but more rare now)
- decide what is good enough to ship
That is a very different job from typing code manually all day. Looks more like team lead role.
We are already seeing the first signs of that adaptation. Stack Overflow notes that 35% of developers now arrive there because of AI-related issues, effectively using human-verified knowledge to clean up machine-generated mistakes. That is not a sign the model failed. That is a sign the ecosystem is reorganising around the new workflow.
The strongest engineers in this new world will not be the ones who proudly refuse agents. They will be the ones who know how to use them properly, when to constrain them, when to verify them, and when to throw their output in the bin.
Manual coding will not disappear, just like assembly did not disappear. But it will become more niche. The future default is “developers who can direct agents replace developers who only type code.”
