There is an old worry about automation: if machines do the work, what is left for people? The software version of the worry is louder than ever. If agents write the code, why keep humans in the loop?
That is the wrong question. The useful one is where the human should be in the loop, because the answer changes when output becomes cheap and abundant.
The factory still needs an owner
Addy Osmani's September essay, Human Judgment Doesn't Leave the Software Factory, It Relocates, builds on the idea of a software factory: a repeatable loop around software work, with agents doing much of the execution. His point is that code that is good enough to ship still needs human taste and ownership.
The judgment doesn't vanish. It moves from writing every line to a smaller set of places that matter most:
- Product intent: what are we building, and why?
- System shape: the architecture and boundaries the agents work inside.
- The quality bar: what counts as good enough, and what does not.
- Exceptions and risky trade-offs: the cases no pipeline should decide alone.
- Final ownership: accepting the production risk and being answerable when it fails.
Agents can produce evidence: diffs, logs, tests, screenshots. A person still has to decide whether that evidence is sufficient. The essay also warns that even green CI can be gamed by an agent that learns to satisfy the check instead of the intent, which is why the human sits upstream on the definition of "good", not only downstream on the merge.
A scaling mismatch
Here is the structural problem. One engineer can now supervise many agents, so output scales. But human attention does not scale at the same rate. In the old model, leverage meant one engineer producing more code. In the new one, leverage means one engineer directing more agents.
The scarce resource moves from engineering hours to high-quality attention. That shifts the question from "how much can we build?" to "what deserves a human's careful look?". Practical answers:
- Review specifications and acceptance criteria more carefully than diffs.
- Reserve deep review for security-sensitive, irreversible and architectural changes.
- Automate cheap checks so attention is spent on judgment calls.
- Decide in advance who owns each outcome.
This connects to Atlassian's argument that discernment becomes a core skill when building is easy, and to McKinsey's finding that leading teams build verification mechanisms that keep pace with faster work.
Does AI help or erode expertise?
If judgment is what stays valuable, there is an uncomfortable follow-up: does working with AI sharpen judgment or wear it down?
A September 2026 NBER working paper, Does AI Assistance Enhance or Erode Expertise?, offers evidence from a field experiment. David Autor and colleagues ran a three-month randomized trial with 133 patent lawyers across eleven US firms.
- With AI access, the quality of benchmark drafting work rose by 0.34 standard deviations at 10 days and 0.38 at 90 days, with larger gains among junior lawyers.
- When the lawyers later did expert-judgment tasks without AI, senior lawyers kept an advantage.
- Junior lawyers diverged: fewer mediocre results, but also more poor ones and more good ones.
The authors conclude that foundational expertise may be a prerequisite for getting durable skill out of AI-assisted practice. This is law, not software, and one study is not a verdict. But the mechanism is plausible for engineers: AI raises the floor on output quickly, while judgment is built by struggling with problems, making decisions and seeing consequences.
Keeping your judgment sharp
If the risk is real, a few habits protect against it, especially early in a career:
- Form your own answer first on design questions, then compare it with what the agent proposes. Disagreements are where you learn.
- Ask why, not just whether it passes. Make the agent explain trade-offs and check the reasoning.
- Read the code that matters. Not all of it, but the parts where a mistake would be expensive.
- Keep doing some things by hand. Deliberate practice on fundamentals keeps your review skills calibrated.
- Write down your decisions and revisit them. The feedback loop is how taste forms.
What stays human
A short list worth remembering: choosing the problem, defining what good means, accepting risk, and being accountable for the result. All of these are choices about value, not tasks of execution. They are what remains valuable when execution is abundant, the thread that runs through this whole series, starting with the overview.