As AI agents take on more of the typing, the obvious question is what developers are left to do. The honest answer is that the role is not disappearing — it is moving up a level, towards the work that has always mattered most and that agents are least able to do alone.
From author to architect
When generating a first version of a feature is fast, the scarce skill becomes deciding what to build and how it should fit together. System design, data modelling, choosing boundaries and trade-offs — this is where good software is won or lost, and it is increasingly where a developer's time is best spent.
From typing to reviewing
Code written by an agent still has to be understood, checked and owned. Reviewing generated code — for correctness, security, performance and fit with the rest of the system — becomes a core daily activity rather than an occasional one. Accepting output you do not understand is how problems reach production.
Supervising agents
Working with agents well means defining what they can do: the scope of a task, the tools they may use, and the point at which a human must approve an action. Good supervision is less about writing prompts and more about setting boundaries and checking results.
Judgement is the constant
Across all of this, the thread that does not change is judgement — knowing what good looks like, where the risks are, and when to stop and think. Tools have changed what developers produce and how quickly; they have not changed the need for someone accountable for the result.
What teams should build
For teams adopting these tools, the payoff comes from more than access to an agent. It comes from engineering standards, review practices and governance that let people move quickly without losing control — and from being clear about who is responsible when an agent is involved.