AI Harness Engineering: Runtime Substrate Identifies 11 Critical Responsibilities for Reliable Agents
arXiv·high signal
Argues that autonomous software-engineering agent reliability is a model-harness-environment system problem, not just a model capability gap. Identifies eleven component responsibilities a runtime substrate must provide: task specification, context selection, tool access, project memory, task state, observability, failure attribution, verification, permissions, entropy auditing, and intervention recording. Proposes a four-level harness ladder (H0-H3) that progressively exposes runtime support to foundation-model agents.