A mental model for building trust in coding agents by constructing an 'outer harness' of feedforward guides and feedback sensors. Guides steer agents before they act (e.g., AGENTS.md, code mods, skills), while sensors detect issues after the fact and enable self-correction (e.g., linters, structural tests, AI code review). Controls are classified as computational (deterministic, fast) or inferential (semantic, AI-based). The model covers three harness categories: maintainability, architecture fitness, and behavioral correctness. The human's role shifts to iteratively steering the harness itself, directing human judgment where it matters most. Open challenges include harness coherence, behavioral verification, and tooling to manage guides and sensors as a unified system.
Table of contents
Metaphors only go so farFeedforward and FeedbackComputational vs InferentialHow does harness engineering relate to context engineering?The steering loopTiming: Keep quality leftRegulation categoriesHarnessabilityAmbient affordancesHarness templatesAshby's LawThe role of the humanA starting point - and open questionsSort: