Failure-driven harness optimization is an iterative development process that systematically converts postmortem findings from agent execution failures into versioned runtime controls without modifying underlying model weights. Following a failed run, defects are categorized into operational causes—such as missing context, invalid transitions, weak verification checks, premature handoffs, or ineffective recovery. The executing agent or a supervisory hyper-agent then proposes targeted modifications to state boundaries, prompts, entry/exit hooks, pre-commit checks, recovery rules, and practice activation criteria. Proposed changes are reviewed, regression-tested against existing tasks, and merged into a new versioned runbook.