Contents

Software engineers should prepare for work that places more weight on choosing problems, shaping the development process, and taking responsibility for outcomes than on writing code. Mori Atsushi’s essay “What I Think Now as a Software Engineer,” published on November 15, 2026, makes that case while acknowledging the loss it entails. The practical next step may look nothing like the career you wanted.

Key Takeaways#

  • Mori’s assessment that Opus now exceeds his abilities is a personal judgment, not proof that every software engineer can be replaced.
  • Invest selectively in AI development workflows, prioritizing improvements that solve measurable problems today.
  • Treat review misses as evidence of insufficient context or verification, while retaining explicit approval for changes with serious consequences.
  • Test new directions through small projects before making major career decisions.

What Mori’s Argument Changes#

Mori describes a shift from treating AI as fast but unreliable coding assistance to trusting it with important engineering decisions. He says Opus finds problems he would probably miss and proposes designs he considers better than his own. AI still makes mistakes. What has changed is his assessment of its overall capabilities.

That distinction matters. An imperfect tool can still change where people should focus their attention. If generating implementations and comparing designs becomes cheaper, engineers need to reconsider where their time creates the most value.

The essay does not establish how broadly this experience applies. It offers no controlled comparisons across models, repositories, or development teams. Read it as a candid account of one engineer’s changing expectations, and test those expectations against your own work.

Build Only the AI Support That Helps#

Mori identifies two immediate responsibilities: creating conditions in which AI can work effectively, and catching dangerous changes. Examples include recording project decisions, giving agents clear implementation guidance, and making verification easier.

His caution concerns how to invest. General-purpose development support may eventually become standard in tools such as Claude and Codex. That is a prediction, not a reason to postpone improvements that already save time.

Start with these four steps.

1. Collect build, test, and formatting commands in one project document and keep it updated. 2. Record the reasons behind important constraints, such as compatibility guarantees and data handling rules. 3. Provide reproducible steps that let agents run the application and check the results. 4. Measure review time and rework for two weeks before adding more automation.

A small amount of well-maintained guidance is easier to evaluate than a large internal framework. Expand it when recurring failures reveal a concrete need.

Turn Review Misses into Improvements#

When an AI-generated change causes a problem, “review more carefully” is not enough. Identify what would have made the problem detectable. It might be a requirement, a reproducible test, access to relevant context, or a clearer scope of work.

For example, if a change breaks a compatibility guarantee, document that guarantee and add checks where possible. If an unstable development environment prevents an agent from verifying behavior, fixing the environment may help more than repeatedly revising its output.

Mori expects the human role in approving changes to shrink as AI gains better context and support. But the essay does not prove that every review miss stems from an inadequate environment. Retain explicit approval for changes with serious consequences, such as destructive data operations or permission changes, and evaluate automation against actual outcomes.

Try Other Ways to Create Value#

Mori is considering deeper involvement in product work, work on AI development infrastructure, and leaving software altogether. Independently releasing an application is also one way to take responsibility for product decisions.

Each option calls for a different experiment. For product work, choose one user problem and define a measurable outcome before development. For infrastructure work, improve one recurring development bottleneck and measure the effect. If you are considering another field, investigate the actual work and entry requirements before committing to a career change.

Try it for four weeks, with a defined deliverable and a review date. At the end, ask two questions. Did the work create value? And do you want to continue? Neither answer guarantees a lasting career, but both are more useful than choosing a path solely because it seems harder to automate.

Separate Loss from Prediction#

The essay is most compelling in its personal passages. Mori can identify rational steps to take while feeling that the career he aspired to is disappearing. Knowing what to do does not erase that sense of loss.

Avoid turning either optimism or discouragement into certainty about employment. The evidence presented does not tell us how much engineering work will remain, which specialties will retain demand, or when current ways of working will change.

Separate observations from predictions. “AI completed this task with less rework” is a testable observation. “My profession has no future” is a much broader conclusion. Base career decisions on repeated evidence, financial constraints, and the work you want to take responsibility for.

The Next Step#

This week, choose one completed development task and write a one-page account of what AI handled, where human intervention mattered, and one change to the process you will try on the next similar task.

Source: https://note.com/at_sushi_at/n/n2e6ea3453e37

Contents