AI Coding Agents Are Changing Software Engineering, Not Eliminating It
Summary
The article argues that predictions of coding being fully solved or programmers disappearing have arrived ahead of reality, even though the underlying direction of change is real. Developers can now describe a feature and receive a pull request with code and tests, while coding agents can handle larger units of work than autocomplete or individual functions. In one Cockroach Labs experiment, a multi-agent pipeline merged more than 1,200 pull requests with seven reverts over five months and added IBM Db2 support in under two days for about $4,000 in token costs; humans retained final approval, which became mandatory after an initial autonomous week. The author says agents' main failures are increasingly mismatches with team intent, standards, scalability, or operational needs rather than obvious syntax errors. Models can produce code that compiles and demos successfully while leaving unresolved behavior during outages, duplicate data, or heavy load, and developers must often debug code they did not write. Labor-market figures cited in the article show software-development postings rising nearly 15% since early 2025 even as overall U.S. postings declined, but senior roles accounted for 71% of that increase and employment among developers aged 22 to 25 fell nearly 20% since 2024. The author is especially concerned that disappearing junior tasks could weaken the path to senior expertise. The emerging role is described as orchestrating agent fleets: defining problems, setting constraints, evaluating models, reviewing output, deploying systems, and taking responsibility for results. Coinbase reportedly found that a small open model trained on its fraud data beat a leading frontier model on its benchmark with less than half the latency, showing why model selection and evaluation are becoming ordinary engineering work. The recommended response is daily practical use, stronger fundamentals in design, security, concurrency, and performance, real-case test sets, closer business understanding, clearer design documents, and end-to-end ownership. The article concludes that software work is shrinking toward judgment, communication, and responsibility rather than disappearing on the timelines commonly predicted.