The market has converged on the vocabulary (intent engineering, harness engineering, context as the moat) but remains structurally divided on the solution. Three independent sources arrived at the same four-level maturity pyramid: prompt, context, intent, specification engineering. The same insight from an arXiv paper, a Forbes framework, and an enterprise consulting series, all landing the same week.
What moved: 65% of enterprise AI failures trace to harness defects (context drift, schema misalignment, state degradation), not model capability. Harness engineering drives a 10x economic differential through KV-cache locality, and production environments are rejecting unconstrained swarms for bounded supervisor patterns with human-on-the-loop circuit breakers. The harness is commoditizing into an enterprise AI control plane: CI/CD plus IAM plus observability for autonomous agents.
The sharpest articulation comes from Jonathan Gardner: “The distinction between AI that fails and AI that succeeds at the wrong thing is the most important unsolved problem in enterprise AI right now.” He proposes Agent Actionable Objectives, machine-readable expressions of what the organization wants agents to optimize for. Nobody owns the translation from strategic intent to agent-actionable objectives. Executives don’t build agents. Engineers don’t see strategy. Alignment is assumed but never verified.
The architecturally concrete piece: value stream maps recast as active control planes, not documentation. Capability models become readiness heatmaps. Policy-as-code in the API gateway lets you govern 10,000 agents without reviewing a single line of code. Autonomy zones for low-ambiguity tasks. Halt-on-exception at 90% confidence. Agents treated as digital insiders with RBAC tied to parent capability.
The dangerous gap: governance and safety remain the least mature layer. 98% of assessed agents exhibit the lethal trifecta (private data plus untrusted content plus external communication). 54% of deployments hit significant incidents within 90 days. Observability deployments are actually declining as adoption outpaces controls.
The Klarna case is the canonical warning: $60M saved, 2.3M conversations in month one, then the CEO admitted cost was “too predominant evaluation factor.” AI optimized for deflection when the business needed retention. By 2026, restructured to freelance human support. The system worked. It optimized for the wrong thing.
The market is building faster than it is governing. The intent layer is the strategic differentiator. Governance is the dangerous laggard.