What we cannot automate yet in developer workflows
AI tools can automate more development work every month, but key parts of engineering still resist full automation for structural reasons.
By Agent Software

Some tasks are automatable, some responsibilities are not
AI-assisted development keeps getting better at bounded tasks: drafting code, summarizing changes, identifying likely files, or proposing tests. That progress is real. The mistake is to map that task-level improvement directly onto end-to-end engineering responsibility.
Engineering is not just the sum of local actions. It also includes judgment about tradeoffs, risk, timing, and organizational context.
Ambiguity is still expensive
The hardest parts of software work are often the parts with incomplete goals or conflicting constraints. Should the system preserve backward compatibility or simplify the API? Is a performance regression acceptable given a release deadline? Does the quick fix create operational debt that will hurt the team later?
These questions are not impossible for AI tools to contribute to, but they are not straightforwardly automatable either.
Verification remains a bottleneck
Automation also hits a limit when verification is weak. If the system can act faster than the team can confirm whether those actions are correct, supervision cost rises instead of falls. That is why evaluation and execution visibility matter so much in the suite story.
Without them, more automation can simply produce more plausible mistakes per hour.
Human trust is part of the workflow
There is also a social boundary. Teams need to understand why a change was made and whether they are comfortable owning it. That requirement does not disappear because a model can produce code quickly.
A good local-first tool stack should acknowledge that reality rather than pretending all friction is irrational.
The near-term answer
We can automate more of the mechanics while keeping the human responsible for interpretation, tradeoffs, and final accountability. That is a more useful framing than the false choice between full autonomy and no progress.

