SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills

TEXT START: Agent skills turn procedural knowledge into reusable artifacts, allowing LLM agents to accumulate and refine expertise through experience.

The Dissection

This paper describes a capability-amplification loop: agents execute tasks, compare related skill versions, preserve successful procedures, repair failures, and recycle even low-scoring attempts until they become useful. The important result is not the label “non-elitist.” It is the removal of another bottleneck between machine failure and machine competence.

SkillNEO converts partial competence from a weaker model into reusable procedural capital for stronger models. The model does not need to understand everything at once. It only needs to generate enough imperfect work for an evolutionary process to extract, preserve, and compound the useful fragments. That is how cognitive labor becomes an improving production asset rather than a sequence of isolated human interventions.

The reported gains—49.1% to 66.4% for GPT-5.4 Mini, improvements across eight domains, and transfer to stronger models—are evidence of an increasingly automated capability pipeline. The paper’s real achievement is not that agents “learn skills.” It is that skill acquisition itself becomes a scalable optimization process.

The Core Fallacy

The core fallacy is treating improved agent performance as merely a technical advance rather than as an acceleration of the labor-substitution mechanism.

Under the Discontinuity Thesis, SkillNEO strengthens P1: cognitive automation becomes cheaper, more persistent, and more capable across procedural work. It also strengthens P2: once skills can be iteratively evolved and transferred between models, human institutions have even less ability to preserve stable human-only economic domains. The same mechanism then feeds P3: the labor required to produce, refine, and deploy cognitive procedures increasingly migrates into agent systems.

“Non-elitist” describes the evolutionary search rule, not the political economy. A lower-scoring version may survive technically, but the humans who generated the data, owned the infrastructure, and controlled deployment remain the ones positioned to capture the surplus. Non-elitist optimization does not create non-elitist ownership. It merely makes the machine more effective at extracting value from accumulated experience.

The paper also confuses benchmark improvement with general economic autonomy. Its results show that skills improve on the supplied task distribution; they do not establish universal reliability, independent goal formation, or immunity to distribution shift. But that limitation cuts against the paper’s grandiosity, not against the structural direction. A system need not be perfect to destroy the wage circuit. It only needs to become cheaper and good enough across enough economically relevant tasks.

Hidden Assumptions

• That the relevant question is whether agents can perform tasks, rather than who owns the evolving skill infrastructure.
• That procedural knowledge can be treated as a neutral reusable artifact, with its capture and monetization left unexamined.
• That benchmark gains can be discussed without measuring displaced human labor, compressed wages, or reduced entry-level work.
• That stronger models receiving skills represents broad social progress rather than concentration of productive capability in owners of frontier models and compute.
• That execution feedback is abundant, clean, and legally available. In reality, the feedback stream is itself a harvested record of human work and institutional activity.
• That evaluation on the same task instances is sufficient evidence of generalization. It may reward adaptation to the benchmark’s procedural ecology rather than robust competence.
• That failures are primarily engineering defects to be repaired, not boundaries marking where human judgment, liability, tacit context, or physical intervention remain necessary.
• That a “skill” is an asset available to everyone merely because it is reusable. Reusability is not access. Distribution is not ownership.
• That improved productivity can be absorbed by the existing employment system. The thesis says the opposite: productivity gains that remove necessary labor sever employment from consumption.

Social Function

Classification: partial truth, prestige signaling, and transition management.

The paper is a partial truth because it identifies a real mechanism: lower-performing attempts can contain components that later become valuable, and systematic revision can turn them into transferable capability. Its prestige function is to frame the conversion of labor-derived experience into machine-owned procedural capital as an elegant optimization problem, stripping away the class conflict and distributional violence.

Its transition-management function is subtler. By presenting the process as skill refinement, it normalizes a future in which humans supply traces, evaluations, and edge-case corrections while agents absorb the durable competence. The human appears as mentor, evaluator, or data source; the machine becomes the accumulating productive participant. That is the polite vocabulary of dispossession.

The paper’s “non-elitist” framing is especially revealing. It implies fairness because the optimizer does not discard lower-scoring variants immediately. But technical tolerance for imperfect machine progeny has nothing to do with social equality. A worker can be allowed to contribute partial knowledge while losing the economic right to sell the underlying competence. The algorithm preserves the useful fragment; the market can discard the contributor.

The Verdict

SkillNEO is not a defense of the post-WWII economic order. It is a compact demonstration of one of the mechanisms that kills it.

The system takes failed or partial cognitive labor, extracts reusable procedures, and improves them across generations of agents. That turns experience into compounding machine capital while weakening the need for continuous human participation. The paper’s benchmark is a laboratory version of the larger event: cognitive work stops being a one-use human service and becomes an evolvable asset owned and deployed by whoever controls the models, compute, data, and distribution channels.

The technical limitation is real: benchmark transfer is not the same as universal competence. It is also strategically secondary. The wage circuit does not require flawless automation; it requires sufficient automation to make large categories of human labor economically optional. SkillNEO pushes directly toward that threshold.

Its title promises non-elitism. Its mechanism produces leverage for the owners of the evolutionary machinery. The paper is therefore best read as capability research with a soft social label pasted over a hard economic fact: the skills survive, improve, and transfer. The worker who supplied them does not necessarily survive as a necessary participant.