Enterprise agent demos keep running beautifully, then stall the second they hit real operations. They execute for a bit, then need a human to top up context and check output. The promised efficiency drains into supervision. The agents did the work, and you did the watching. That is why so many agent pilots never become production systems.

One of the more concrete attempts to break that pattern comes from Nace.AI. The company says its agents handle the bulk of regulated workflow while human experts validate the result, marketing a 90/10 split. Nace uses a generator it calls a MetaModel, which produces parameter adaptations for a model at inference time from a company's policies. The point is to keep the system competent longer without relying on either ever-growing prompts or full retraining cycles.

To understand why this matters, start with the uncomfortable truth about how agents “forget” or “misread” as they run. Fine-tuning bakes business knowledge into a model’s weights, but it can still suffer catastrophic forgetting, a problem identified in the 1980s and still unresolved in 2026. Teams often work around it by isolating each task in its own fine-tuned model or adapter. That produces a sprawling model estate, raising cost and governance overhead. And because a fine-tuned model is essentially a snapshot, it can go stale the day a policy changes, forcing an expensive, slow retraining cycle.

In-context learning (often paired with RAG) tries to avoid retraining by supplying relevant policies at run time. But this is where context rot bites. Retrieval narrows what goes into the prompt, yet a retrieval miss can look identical to a confident answer. Meanwhile, the mechanics of stuffing more tokens into the prompt push cost and latency upward as context grows. The failure modes rhyme across the two approaches. With fine-tuning, an agent can be confidently using last quarter’s policy. With in-context learning, it can be confidently using a detail it lost in the middle of a long prompt. Either way, the output can look equally assured, so you cannot reliably tell what is wrong without checking everything. That is the reason the human never truly leaves.

Some teams run both approaches at once: fine-tuning stable knowledge and retrieving the rest. This “softens” each failure, but does not eliminate the underlying uncertainty. You still cannot be sure the model is both current and using the right context for a given answer, which keeps review requirements in place. The deeper problem is not simply model capability. It is where your knowledge sits relative to the model, and what happens to that relationship as the task unfolds.

Nace’s third path is designed to move that relationship closer to the moment of execution. Instead of retraining one model or stuffing a prompt, a generator builds a small, task-specific model on demand from company policies at inference time. This generator is a hypernetwork, described as a network whose output is the weights of another network. The idea was named in 2016, and applying it to produce specialist language models from text or documents is recent and active. Sakana AI’s Text-to-LoRA, presented at ICML 2025, generates a model adapter from a plain-language description in a single pass. A 2026 system called SHINE calls hypernetwork adaptation a promising new frontier, precisely because it sidesteps both retraining cost and context limits of prompting.

Why collapse the library? Hypernetwork-generated adapters are meant to replace a sprawling set of per-task LoRAs with one network that can produce adapters on demand, including for tasks it has not seen before. The elegant part is that the thing teams build manually to dodge catastrophic forgetting becomes the same object the hypernetwork produces automatically. In theory, that reduces governance headaches by treating “model specialization” as generated output, not as an ever-growing zoo.

This also matches the case for going small inside agent workflows. A 2025 paper by Nvidia researchers argues that for narrow, repetitive tasks, small models can be capable enough and 10 to 30 times cheaper to run than frontier generalists. Nace’s focus is on regulated, pointed use cases like audit, compliance, and risk assessment. The company markets a division of labor where agents do most of the workflow and human experts validate the result, aiming at 90/10. But the autonomy claim is not just a number; it is a property of architecture. Two design choices determine whether that autonomy is trustworthy or merely fast.

First is grounding, tying outputs to their source so reviewers can verify rather than redo. Research models like HalluGuard label each claim as supported or not and cite the passage they relied on. Nace ships its agents with grounding models and reasoning traces for the same reason. A “10% review” only means something if the human can confirm provenance in seconds.

Second is the feedback loop: when experts validate outputs, whose model improves, and where does it live? That decides whether the compounding asset belongs to the vendor or to you. Arrangements differ. Nace, for instance, uses an external network of certified experts for some engagements, and for direct enterprise deployments, it uses the customer’s own staff, with the resulting model kept inside the customer’s cloud. In other words, ownership and learning routing are part of the product design, not an afterthought.

Still, this third path is early, and a few unanswered questions will decide whether it scales beyond pilots. Calibration is the linchpin. The approach depends on the model knowing when it is unsure, because uncertainty is what triggers escalation. Recent work generating these adapters found they do not automatically improve calibration over ordinary fine-tuning, with gains appearing only under specific constraints. Quality also depends heavily on the policy data the system is built from, which raises the premium on data curation. And scale remains an open research frontier, since published hypernetwork examples so far have been small. If calibration does not improve and data quality wobbles, the “90/10” promise becomes just another demo-friendly story.

For executives evaluating agent readiness, the strategic stakes are simple: autonomy is not a dial you set. It is an outcome of how your system anchors knowledge, manages uncertainty, and composes learning ownership. Hypernetwork-generated task models may be a credible way to raise the autonomy ceiling, but only if grounding is fast, calibration is honest, and the feedback loop compounds inside the right organizational boundary.