Ontological Reflections on Agent and Skill

I. Introduction: A Foundational Question That Cannot Be Answered

“What is an Agent? What is a Skill?”—this is the first question anyone entering the field of LLM application development encounters. Yet, after systematic research across the official documentation, academic papers, and engineering practices of the three major AI platforms, we have arrived at a disquieting conclusion: these two concepts currently have no rigorous definitions.

This is not because the definitions are too specialized or too technical, but because the definitions themselves are ambiguous, self-contradictory, and vary from platform to platform. The deeper issue is that this ambiguity is not an accidental oversight—it is a structural predicament inevitably encountered when an industry built on probabilistic systems attempts to describe itself in deterministic language.

Starting from the author’s own practice as a deep LLM user, this paper reverse-traces every link in this “chain of ambiguity”—from concept definitions to categorical boundaries, from cybernetic ceilings to cost black boxes, from folk complaints to market structure—ultimately revealing a complete causal loop.

II. The Ontology of Agent: The Emptiness of Definition

2.1 Cross-Comparison of Three Definitions

The three major global LLM platforms currently define Agent differently, and they contradict each other on key dimensions:

All three are ambiguous, but ambiguous in different ways. OpenAI is the most pragmatic^[2]—an Agent is simply “a configured LLM instance”; Google is the most expansive^[3]—virtually any AI system with reasoning capability qualifies; Anthropic is the only one that attempts to draw an architectural boundary (Workflow vs. Agent)^[1], but this line itself is unclear.

2.2 Agent vs. Workflow: The Only Boundary Line and Its Ambiguity

Among the three, Anthropic is the only one that attempts to draw an internal boundary. Anthropic explicitly distinguishes: a Workflow is a system that orchestrates LLMs and tools through predefined code paths; an Agent is a system where the LLM dynamically directs its own processes and tool usage. The criterion for this boundary is the locus of control—who decides what to do next? If it is predetermined by code, it is a Workflow; if the LLM decides for itself, it is an Agent.

OpenAI and LangChain also recognize this distinction. LangChain notes that both OpenAI and Anthropic treat Workflows as a design pattern distinct from Agents—in Workflows, the LLM has less control and the process is more deterministic. However, both also acknowledge that for many applications, Workflows provide predictability and consistency for well-defined tasks, while Agents are better suited for scenarios requiring flexibility and model-driven decision-making. In other words, this is not a hierarchy of superiority, but a matter of appropriate use case.

The problem is that this boundary in practice is a continuous spectrum, not a discrete dividing line:

Most of what people call Agents are not actually Agents. A large number of so-called “Agents” are merely an API call plus tool access—they cannot act independently or make decisions. They simply respond to the user. Yet we still call them Agents. An independent researcher pointed out that CrewAI’s so-called “Agents” are actually closer to predefined Workflows assigned to specific tasks, whereas Anthropic’s definition of an Agent is a system capable of independently reasoning through any task^[25]. Both interpretations have their value, but they point to fundamentally different technical entities.

More critically, for most applications, production-grade agentic systems will be a combination of Workflows and Agents—pure Agents rarely appear in production environments. Anthropic recommends finding the simplest viable solution and only adding complexity when necessary. This may mean not building an agentic system at all. This amounts to saying: the concept of Agent, despite being defined, is in practice recommended to be avoided as much as possible. A concept whose own definers advise cautious use has questionable validity as a product category.

2.3 The Paradox of Autonomy

The core of the Agent definition lies in “autonomy.” But deeper interrogation reveals a fundamental paradox: an Agent without a prompt has no autonomy.

The “autonomy” of an LLM is not intrinsic but prompt-conferred. A bare LLM—without a system prompt, without role definition, without tool descriptions—does only one thing: predict the next token based on input text. It will not proactively set goals, decide to call tools, or judge whether a task is complete.

2.4 The Deep Causes of Definitional Ambiguity

Anthropic itself acknowledges: “Agent can be defined in several ways.”^[1] This is not a rigorous ontological definition; this is a product classification strategy. The ambiguity has three causes: first, in 2026, “Agent” is the hottest keyword—the wider the definition, the larger the market narrative; second, the autonomy of Agents is indeed a continuous spectrum, and drawing a hard line is meaningless in engineering terms; third, the field is so new that even academia is still debating.

III. The Ontology of Skill: Conceptual Drift

3.1 What Skill Is Not

The essence of Skill must be delineated through negative definition:

A Skill is not a Tool—a tool is a deterministic operation: give it input, get structured output. A Skill is a set of instructions interpreted by the LLM.^[4] A Skill describes “how to do” something; a tool “executes” something.

A Skill is not a Prompt—a prompt is ephemeral, reactive, embedded in code. A Skill is a persistent, portable, version-controlled artifact.^[5]

A Skill is not an Agent—an Agent is an execution runtime with its own tools, memory, and decision loop. A Skill is a knowledge module that any Agent can load.

3.2 The Three-Way Split

Microsoft’s Semantic Kernel initially called them “Skills,” then renamed them to “Plugins”^[6]—indicating that in their view, there was never an essential difference between Skill and Tool. This directly challenges the thesis that “Skill and Tool are different categories.”

3.3 Self-Examination and Collapse of the Tool–Skill Boundary

The common assertion that “Tool is the execution layer, Skill is the orchestration layer” was subjected to four-dimensional verification (logic, definition, category, analogy), and every dimension exposed problems:

The most honest conclusion: the difference between Tool and Skill is not two essentially distinct categories, but different positions on the same continuous spectrum.^[7] The industry itself has not yet reached consensus.

IV. The Cybernetic Perspective: The 95% Ceiling and the Duality of the 5%

4.1 Quantifying Uncertainty

Through long-term observation of prompt reusability, output stability in Project environments, and Skill reuse effectiveness, all experience converges toward the same direction: there exists an approximate 95% empirical ceiling on the controllability of LLM output. This is not a precise constant measured through controlled experiments, but a convergent judgment based on sustained practice—interestingly, MIT’s independent research also found that 95% of generative AI pilot projects fail to reach production^[21], and these two “95%” figures across different dimensions form a thought-provoking resonance.

This final 5% is not an engineering problem—it is not about prompts being poorly written, Skills being insufficiently refined, or the wrong framework being chosen. It is a physics-law-level limitation determined by the very mechanism of LLMs generating tokens through probabilistic sampling.

4.2 The Value Inversion of the 5%

The key insight is this: the 5% of uncontrollability is a risk from the cybernetic perspective, but the core value from the creativity perspective.

When an LLM deviates from expectations—if the deviation is “worse,” we call it hallucination or error; if the deviation is “better,” we call it brilliance or insight. Both originate from probabilistic sampling producing unanticipated output, but they are not strictly equivalent—academic research shows that different hallucination-suppression techniques have opposite effects on creativity: some methods (such as CoVe) actually enhance creative diversity while reducing hallucination, while others (such as DoLa) suppress both simultaneously^[22]. Across different decoding layers of a model, there exists a quantifiable hallucination-creativity tradeoff curve, with a specific optimal balance point^[23].

4.3 LLM as a Novel Product Category

This 95%+5% structure defines a product category unprecedented in the history of human commerce:

No existing pricing model, evaluation framework, or quality standard was designed for this kind of product. This is the root cause of why every existing concept fails when applied.

V. Folk Ontology: “Gacha” and “Dumbing Down”

5.1 “Gacha”—The Most Honest Description of Cost

“Gacha” is not a term coined by any expert. It is a consensus vocabulary that hundreds of millions of Chinese-speaking AI users spontaneously converged upon through real-world usage^[18]—originating from gaming communities, where users have long understood the nature of probabilistic systems: spending money guarantees nothing, rarity determines cost, and operators can secretly alter the odds.

“Gacha” reveals a fact ignored by pricing models: the real cost of production is not “the price of a single generation,” but “the total cost of pulling until you get a satisfactory result.” And this total cost is unpredictable.

5.2 “Dumbing Down”—Experiential Convergence of Three Distinct Mechanisms

What practitioners observe as “dumbing down” is actually the convergence of three different mechanisms at the level of user experience:

Mechanism One: Data Quality Degradation (“Brain Rot”). Research from multiple universities has shown that models trained on low-quality internet data exhibit systematic capability decline^[8]—reasoning scores dropped from 74.9 to 57.2, and memory and long-context comprehension fell from 84.4 to 52.3. More alarmingly, this damage is difficult to reverse—even retraining with high-quality data cannot fully restore the model to its original level. This is a data-source-level problem.

Mechanism Two: The Alignment Tax. RLHF-aligned models exhibit “response homogenization”—research has measured that on TruthfulQA, 40% of questions produce only a single semantic cluster across 10 independent samples^[24]. For affected questions, sampling-based uncertainty estimation methods see their discriminative power drop to zero (AUROC=0.500). This is not capability degradation but a structural price paid for safety—a design choice.

Mechanism Three: Cost Optimization and Model Substitution. Vendors perform cost-optimization inference tuning on newer models, defaulting to shorter outputs; simultaneously, they may silently replace high-cost models with low-cost ones while keeping billing unchanged^[9]. This is a profit-driven business decision.

The three mechanisms have different causes, different accountability, and different resolution paths. But at the user experience level, they converge into the same feeling: what worked before no longer works; what used to be brilliant responses have now become mediocre. The folk term “dumbing down” precisely encodes this composite experience—even though it does not distinguish the causes.

5.3 The Opposition of Two Language Systems

Officially, the industry speaks of Agent, Skill, Tool; users speak of gacha and dumbing down. Two language systems describe the same system, yet arrive at entirely different conclusions. The former says this is a controllable, definable, priceable product; the latter says this is an uncontrollable, luck-dependent, progressively worsening gamble. The folk vocabulary of “gacha + dumbing down” is far more honest than the official vocabulary of “Agent + Skill.”

VI. The Cost Black Box: Reasoning Tokens and the TPM Trap

6.1 Hidden Reasoning Tokens

Between input and output, there exists a cost layer that has been almost entirely opaque across all companies: Reasoning Tokens. Academic papers have documented cases where over 90% of billed tokens were never displayed to the user, with internal reasoning inflating token usage by more than 20×.^[10]

You send 50 tokens, receive 100 tokens, but are billed for 650 tokens.^[11] Those 500 “reasoning tokens” are the model’s internal monologue—you never see them. Reasoning tokens can amplify the cost of a single query by 5 to 50 times^[12], depending on task difficulty and model selection.

6.2 The TPM Trap: Empirical Data

The following data comes from the author’s (LEECHO Global AI Research Lab) Google Gemini API paid Tier 1 dashboard screenshots^[13], covering a 90-day usage period:

6.3 The Complete Structure of the Quadruple Black Box

Stacking all findings together, the real situation facing an LLM user is a quadruple black box:

Users do not know what quality of result they will get, do not know how much this query cost, do not know why they were suddenly rate-limited, and do not even know whether the model has been silently swapped out behind the scenes^[10].

VII. From Conceptual Ambiguity to Market Black Box: The Complete Causal Chain

All findings linked together form a complete causal chain of reasoning:

Every step in this cycle has already been validated in the real market: approximately 80% of enterprises report using generative AI, but an equal number report no significant bottom-line impact^[14]. Only 11% of organizations have deployed AI Agents to production^[15]. Most enterprise budgets underestimate true total cost of ownership by 40–60%^[16]. MIT further reports that 95% of generative AI pilot projects fail^[21].

A possible counterargument must be addressed: “cloud computing” was similarly ambiguous in its early days, yet it still developed into a well-functioning trillion-dollar market. Why is conceptual ambiguity so much more destructive in the LLM market? The answer is that cloud computing had ambiguity only at the level of conceptual definition, while its output was deterministic—storing 1GB is 1GB, computing for 1 hour is 1 hour, verifiable, auditable, and comparable. The distinctive feature of the LLM market is that ambiguity simultaneously permeates three dimensions: product definition is ambiguous (what is an Agent?), output quality is non-deterministic (the same prompt does not guarantee the same result), and cost measurement is opaque (reasoning tokens are invisible). It is this triple-stacked uncertainty that is the root cause of this market’s inability to form an effective pricing mechanism.

Pricing chaos is the direct symptom: some vendors charge per resolution, some per conversation, and others hide their pricing entirely behind sales calls^[19]. There is not even consensus on billing units—Token, Credit, “Intelligence Unit,” Conversation, Resolution^[20]—none of these billing dimensions represents a value metric from the user’s perspective.

7.1 The Structural Penalty on OOD Users

The most ironic aspect is that the current cost structure systematically penalizes high-value usage and rewards shallow usage. A cross-disciplinary deep thinker—generating high-density OOD (Out-of-Distribution) queries, triggering the model’s dense reasoning mode, continuously stacking context—sees token consumption grow exponentially. These are precisely the users doing what AI should be used for—deep thinking—yet the entire business model penalizes them.

Empirical evidence: the author of this paper, solely through chat (not batch tasks, not Agent automation), triggered Gemini 3 Pro’s 1M TPM limit and was banned for the day. The conversation pattern was characterized by each turn opening a new disciplinary dimension, with context continuously expanding^[17] and reasoning tokens growing exponentially. Without restrictions, this type of conversation pattern could theoretically reach 1B tokens/minute in consumption.

VIII. Conclusion: A New Conceptual Framework Is Needed

8.1 Three Questions the Industry Must Answer

If the market chaos caused by conceptual ambiguity is to be truly resolved, what is needed is a new conceptual framework native to LLM systems—not metaphors borrowed from human organizations, traditional software, or cognitive science, but one derived from the actual operating mechanisms of LLMs. This framework must at minimum answer:

First, the question of atomic units. What is the truly irreducible atomic unit of an LLM system? Is it a token? A single inference call? A complete context window?

Second, the question of isolation boundaries. What guarantees the isolation boundaries between different components? Traditional software has type systems, interface definitions, and process isolation. In LLM systems, everything is mixed in the same token stream—prompts, tool descriptions, Skill instructions, user input, model output—with no true physical isolation, only semantic conventions.

Third, the question of autonomy. What does “autonomy” actually mean in a system that is fundamentally a conditional probability generator?

8.2 Historical Analogy and Outlook

Technology running ahead of concepts is a recurring pattern in the history of technology. When electricity was first invented, it was called “artificial lightning”; when automobiles first appeared, they were called “horseless carriages.” These names all forced old concepts onto new phenomena. “Agent” is borrowed from human representatives and reinforcement learning agents; “Skill” is borrowed from human skills—every one of them is inaccurate.

What will ultimately survive are two kinds of companies: those that have genuinely solved a clear problem that can be evaluated and priced; and infrastructure companies that possess foundational model capabilities and do not depend on the “Agent” narrative. Those in the middle—surviving on ambiguous concepts and capital narratives—will be the first to be eliminated when the market clears.

Only when enough buyers stand up and say “I don’t understand, not because I’m stupid, but because you haven’t explained it clearly—and I’m not buying until you do” will the black-box market be forced to become a transparent market.