Coding agents can read UserService.ts, so instructions shift to context and constraints

The essay frames this as a moving target rather than a fixed best practice. When models could not explore a whole codebase and struggled to run long autonomous tasks, breaking work into steps like opening a file or fixing a method was a rational way to keep the AI from getting lost. Agents that can search, read files and inspect existing patterns change that calculus, because the human no longer has to be the one to decompose the work.

It also draws a line around delegation. Letting the AI choose implementation details does not remove the human role: defining the problem and done conditions, supplying business knowledge that code cannot express, setting constraints and exposure, verifying output, and making the final call on production release stay with people. The essay points to domain distinctions — for example, when customer, applicant and contractHolder are genuinely different business concepts, that difference reflects the domain model rather than naming taste, and the AI may not be able to infer it from code alone.

A caution runs alongside the argument. Detailed instructions can narrow the options the AI considers, so the author suggests letting it investigate and compare implementation approaches. But the wider the delegated range, the more important it is to decide in advance what to check to catch failures, and work without tests, review, diff checks or rollback should not be handed over unconditionally.