Harness Engineering: The Meta Layer as a First-Class Discipline for Multi-Agent Systems

The reliability of an LLM agent depends not only on the underlying model but on the harnessaround it — the loop, tool dispatch, memory, context construction, verification, and governancethat turn a token predictor into an agent. We argue that this harness, and in multi-agent systemsthe governing meta layer specifically, is a separable engineering discipline. We adopt the emergingterm harness engineering and argue it has its own objects, failure modes, metrics, and optimizationsurface. We make three contributions. First, we locate the discipline in the meta layer of a twolayer architecture: a working layer of task experts governed by a meta layer of domain-agnosticprimitives, with scale absorbed by recursion and an immutable constitutional directive rather thanby added hierarchy. Second, we give a taxonomy of six modular, independently ablatable metaprimitives: proactive gap detection, timing-aware escalation, tier-aware calibration, sleep-styleconsolidation, distributed metacognition, and validation-gated self-evolution, with three furtherbrain-suggested primitives (dreaming, incubation, curiosity) sketched as design hypotheses.Third, we distil cross-cutting production principles: a cheap-gate→LLM invariant, a sharp linebetween a harness and a framework, the “domain-agnostic mechanism, domain-specific oracle”formula, and two honest limits (the harness-only ceiling and single-verifier Goodhart). Weposition the discipline against concurrent systems and close with open problems. This is aposition and taxonomy paper: we define the meta layer as an object of study and propose anablation-ready decomposition, separating prior results from our synthesis and from unvalidateddesign hypotheses, and we specify the ablations that would test each.