AI-native dev: drop fixed stages, define goals and ハーネス instead

The essay's starting point is a common design: one agent per traditional stage, chained in order. It grants that this is reasonable for automating an existing process, but argues that splitting stages is not the same as redesigning development for AI. Tasks differ — fixing an outage may start with log investigation, a clear API addition may go straight to implementation and tests.

The piece then separates the target state from the route to it. For a user-registration feature, constraints such as required fields, duplicate-email prevention, and no unintended changes to existing API contracts are defined up front, while class structure and which files to touch are left to the agent after it investigates. This is where the essay positions Spec-Driven Development: useful as a definition of goals, harmful if turned into a fixed sequence for every task.

From there, the piece sorts responsibilities: skills supply conventions and reusable know-how, tools make execution and verification possible, and the harness supplies constraints, feedback, and stopping conditions. The argument is to add only what is missing — a sub-agent for independent review, or external enforcement of permissions — rather than build an orchestrator for its own sake, citing OpenAI's Agents SDK and harness-engineering notes as reference points while noting the design is its own proposal, not a reproduction of an official architecture.