Anthropic Research Pushes Spec-Driven AI Orchestration Over Prompt Engineering

Anthropic's engineering team has published research arguing that the infrastructure surrounding AI models matters as much as the models themselves, coining the term 'harness engineering' to describe this discipline. The research highlights recurring production failures including context degradation, over-ambitious single-pass agents, and unreliable self-evaluation in deployed systems. As teams scale from a handful of workflows to dozens, manually managed scaffolding breaks down and coordination overhead grows nonlinearly across multi-agent setups. The proposed shift moves away from bespoke prompt-based harnesses toward specification-driven contracts that define deliverables, verification criteria, agent assignments, and cost controls in structured but plain-English terms. Predicate Ventures' Blake Aber argues this declarative approach mirrors how Terraform and Kubernetes transformed infrastructure management, potentially compounding efficiency gains across entire product portfolios.