What causes weak models to fail at activating harness artifacts?

This explores why weaker models fail to *activate* harness artifacts — not why they can't read instructions, but why they don't reach for the memory, skills, and protocols a harness offers even when those are available. The cleanest answer in the corpus is that benefiting from a harness is a different skill from producing one, and it follows an inverted-U curve: the capacity to write useful harness edits stays roughly flat across model tiers, but the capacity to actually *invoke and benefit from* them peaks at mid-tier and collapses at both ends — weak models fail to invoke the harness at all, while strong models trip over faithful instruction-following Do stronger models always evolve harnesses better?. So weak-model failure isn't about generating bad scaffolding; it's about not triggering it when the moment calls for it.

Why does invocation specifically break down? Because a harness only pays off when its cognitive burden has been externalized into structure the model is supposed to follow Where does agent capability really come from? — and following that structure depends on the very instruction-following ability weak models are short on. The chain-of-thought research sharpens this: CoT in weaker models is constrained imitation of reasoning *form*, not abstract inference — format matters more than content, and scaling reasoning ability actually creates instruction-following deficits What makes chain-of-thought reasoning fail in language models?. A harness artifact is essentially an instruction-following contract ("when X, consult skill Y; record Z to memory"), so a model that imitates surface form without tracking intent will skip the trigger conditions entirely.

There's a second, sharper failure: even when a weak model produces the right move, it can't reliably *select* it without an external signal. A committee of weak-model calls only matches a strong model when a local soundness check — tests, proofs, type checks — converts latent-correct proposals into actual selections; sampling alone amplifies coverage but can't pick the winner When can weak models match strong model performance?. Read against harnesses, this says weak models need the harness's verification scaffolding precisely *because* they can't self-select — yet activating that scaffolding is the thing they fail at. It's a chicken-and-egg trap: the artifact that would rescue them is the one they don't know to invoke.

The failure also compounds over a trajectory. Once a weak model misfires, its own prior errors fill the context and bias everything downstream — a self-conditioning effect that degrades long-horizon performance non-linearly, and that scaling does *not* fix (only test-time thinking does) Do models fail worse when their own errors fill the context?. The multi-agent literature names the concrete symptoms: role flipping, infinite loops, and conversation deviation, all rooted in LLMs lacking persistent goal and role representation Why do autonomous LLM agents fail in predictable ways?. A harness protocol assumes the agent holds onto a goal long enough to recognize when an artifact applies — and that stable goal-holding is exactly what weak models don't have.

The quietly surprising takeaway: making the model bigger isn't the fix, and at the top it's even counterproductive. Weak models under-invoke harnesses; strong models over-literally follow them or drift from faithful execution — so the sweet spot for harness leverage is the middle, not the frontier Do stronger models always evolve harnesses better?. If you want a weak model to use its scaffolding, the lever is structure that lowers the activation bar — verifiable triggers and external soundness signals — not raw capability Where does agent capability really come from?.