A coding agent can produce a plausible patch with a good prompt. Reliable progress across a real repository needs more than prompt quality.
The Harness Engineering roadmap treats the environment around the agent as a system. The first build turns implicit expectations into scoped repository instructions and executable checks. The second makes the workspace readable from a fresh checkout: structure, setup, readiness, and permissions should be discoverable without private history.
Continuity arrives in the third build. A later session needs the goal, current progress, evidence, blockers, and next action. Without that durable state, long-running work becomes a series of confident restarts.
The fourth build adds scope control and runtime feedback. The agent needs a bounded authority surface and evidence that tells it whether to continue, replan, complete, or escalate. This is where a harness starts to look less like a prompt wrapper and more like an operating system for work.
The fifth build separates maker and checker. The checker reconstructs the contract and gathers fresh evidence instead of accepting the maker's claim that the task is done. The final two builds connect initialization, execution, recovery, verification, and terminal states, then compare the automated loop with a baseline and remove one component to see whether it mattered.
This sequence makes one distinction clear: agent capability and agent reliability are different properties. A strong model can still fail in an unreadable workspace, lose state between sessions, exceed its scope, or verify its own assumptions.
The seven-build roadmap is at https://learn.significanthobbies.com/curriculum/roadmaps/harness-engineering.