Software developer Jake Goldsborough has published an essay arguing that coding agents are changing the composition of software engineering rather than eliminating the profession. His central distinction is between accidental complexity—the syntax, glue code and undocumented commands surrounding a problem—and the judgment required to decide what a system should do and whether a change is safe to ship.
Goldsborough writes that agents can traverse unfamiliar repositories, produce parsers and recall library interfaces faster than a person working manually. Those conveniences reduce time spent translating an understood solution into code. But he also describes the systems making incorrect assumptions, ignoring local conventions, citing nonexistent interfaces and declaring tasks complete without adequate tests. In his account, easier code generation leaves engineers responsible for reproduction, architecture, tradeoffs, verification and production outcomes.
The essay presents the change as a movement of work rather than its disappearance. When an agent writes a data parser, for example, a person still needs to define malformed-input behavior, detect dropped records and understand the operational consequences. Goldsborough argues that this distinction matters because implementation difficulty has often been confused with value. Memorizing a tool’s syntax can be demanding without being the part of a project that most needs professional judgment.
He also acknowledges a training risk. New engineers traditionally develop instincts by struggling through implementation details, and accepting generated changes without inspecting or running them could weaken that learning. His proposed alternative is an iterative loop: ask, inspect, execute, break, read and correct. The essay says agents have helped him explore unfamiliar languages and systems because they lower the startup cost, while the process of checking their output exposes the boundary of his understanding.
Faster generation may increase the importance of engineering controls. A mistaken premise can now produce thousands of lines quickly, so requirements, tests, review and observability become constraints on the volume of plausible-looking output. Goldsborough also flags unresolved industry questions, including companies using productivity claims to justify staffing cuts, generated defects arriving at greater scale, changes to the junior-to-senior career path and unequal access to expensive tools.
The piece is an individual practitioner’s analysis, not an empirical study of employment or productivity. It does not establish how coding agents affect the profession across companies, nor does it dismiss the economic disruption their adoption may cause. Its narrower claim is that generating code and engineering a dependable system are different activities. As tools accelerate the first, responsibility for the second remains with the people who specify, test, review and operate what is built.
