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 harness
around it — the loop, tool dispatch, memory, context construction, verification, and governance
that turn a token predictor into an agent. We argue that this harness, and in multi-agent systems
the governing meta layer specifically, is a separable engineering discipline. We adopt the emerging
term harness engineering and argue it has its own objects, failure modes, metrics, and optimization
surface. We make three contributions. First, we locate the discipline in the meta layer of a two
layer architecture: a working layer of task experts governed by a meta layer of domain-agnostic
primitives, with scale absorbed by recursion and an immutable constitutional directive rather than
by added hierarchy. Second, we give a taxonomy of six modular, independently ablatable meta
primitives: proactive gap detection, timing-aware escalation, tier-aware calibration, sleep-style
consolidation, distributed metacognition, and validation-gated self-evolution, with three further
brain-suggested primitives (dreaming, incubation, curiosity) sketched as design hypotheses.
Third, we distil cross-cutting production principles: a cheap-gate→LLM invariant, a sharp line
between 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). We
position the discipline against concurrent systems and close with open problems. This is a
position and taxonomy paper: we define the meta layer as an object of study and propose an
ablation-ready decomposition, separating prior results from our synthesis and from unvalidated
design hypotheses, and we specify the ablations that would test each.