The diluted word
Every vendor calls their product an "agent" now. Chatbots are agents. API wrappers are agents. Shell scripts that call an LLM are agents. The word has been stretched so thin it covers everything and therefore means nothing. A 2025 AIES paper put it bluntly: "The term 'agent' has been diluted beyond utility and requires redefinition."[1]
This post is our attempt to fix that. We apply the scientific process — observe, hypothesize, test, synthesize — and the 5W framework to reconstruct a rigorous definition of "agent" from first principles. We compare human and AI agency, propose a complexity scoring system, diagram the multi-agent workflows we actually use in this project, and draw a line between what is and is not an agent.
The 5W framework
Before defining what an agent is, we must establish where the definition comes from. The 5W framework — Who, What, Where, When, Why — gives us the axes of analysis.
f: P* → A. [3]The scientific process: three definitions, one synthesis
Instead of picking one definition, we apply the scientific process: examine three independent frameworks, identify their convergent claims, and synthesize.
Hypothesis 1: The Floridi-Sanders observables
Luciano Floridi and J. W. Sanders (2004) proposed that agenthood is not a binary property but depends on the level of abstraction (LoA) at which you observe a system. At a given LoA, an entity is an agent if it exhibits three observables: [2]
The key insight: a thermostat is an agent at a low LoA (it interacts, has autonomous state changes, and adapts to temperature). A human is an agent at a higher LoA (with additional observables like intentionality and moral reasoning). The LoA determines what counts as agency — but the three observables are the minimum.
Hypothesis 2: The Russell-Norvig percept-action function
Stuart Russell and Peter Norvig, in Artificial Intelligence: A Modern Approach (1995, 4th ed. 2020), define an agent as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators." [3]
A rational agent selects the action that maximizes expected performance, given the percept sequence and built-in knowledge. Rationality depends on four things: the performance measure, the percept history, prior knowledge, and available actions. Crucially, rational ≠ omniscient — the agent may lack information. And rational ≠ clairvoyant — outcomes may differ from expectations. This is why rational agents must explore and learn. [3]
Hypothesis 3: The Wooldridge-Jennings weak notion
Michael Wooldridge and Nicholas Jennings (1995) proposed a "weak notion of agency" — the most general, least contentious definition — consisting of four properties: [5]
The "strong notion" adds mental states: beliefs, desires, intentions (the BDI model, inspired by Bratman's 1987 planning theory of intention [6]). The weak notion is sufficient for software agents. The strong notion is needed for philosophical agency.
Synthesis: the convergence
All three frameworks converge on three necessary conditions. We formalize this as:
Where E is the entity and LoA is the level of abstraction. Adaptability (Floridi-Sanders), rationality (Russell-Norvig), and proactiveness (Wooldridge-Jennings) are gradations — they determine the degree of agency, not its presence. A system with all three necessary conditions but no adaptability is still an agent, just a simple one.
Agency vs. autonomy: a critical distinction
The OECD's 2026 report on agentic AI makes an important distinction that most discussions miss: [7]
This distinction maps to our project architecture. Our /build script is autonomous (runs without intervention) but not agentic (has no goals, no adaptation, no planning). Our Cascade agent is both autonomous (operates across sessions without per-action approval) and agentic (formulates sub-goals, adapts strategies, reasons about its own process).
Human agent vs. AI agent: a comparison
With a rigorous definition in hand, we can compare human and AI agents along the dimensions that matter. The comparison is not about capability — it's about the nature of agency.
| Dimension | Human Agent | AI Agent |
|---|---|---|
| Sensors | Eyes, ears, skin, proprioception — multimodal, embodied | Text input, API responses, file contents, network packets — symbolic, disembodied |
| Actuators | Hands, voice, movement — physical world manipulation | Code edits, shell commands, HTTP requests — symbolic world manipulation |
| Goal Formation | Endogenous — desires emerge from biological drives, values, social context | Exogenous — goals are delegated by humans or encoded in system prompts |
| Intentionality | First-order (I want X) and second-order (I want to want X) — Frankfurt's hierarchical volition [8] | First-order only — pursues delegated goals, cannot form second-order desires about its own goals |
| Adaptability | Lifelong learning, transfer across domains, few-shot from single examples | In-context learning within session, fine-tuning across sessions, no cross-domain transfer without retraining |
| Temporal Coherence | Continuous — identity persists across decades, memories consolidate during sleep | Episodic — session-bound, identity reconstructed each session from persistent state (memories, dream graph) |
| Moral Agency | Full — can be held responsible, experiences moral emotions, has free will (debated) | Functional — Floridi-Sanders "mind-less morality": can be evaluated by observable actions without requiring mental states [2] |
| Mortality | Biological — death is irreversible, creates urgency for consolidation | Session-bound — "death" is reversible (new session), but working state is lost, creating analogous consolidation pressure |
| Self-Model | Phenomenal — subjective experience of self, introspective access | Functional — explicit self-model stored in files (our self.md), updated through introspection cycles, no phenomenal experience |
Frankfurt's hierarchy and the AI gap
Harry Frankfurt's 1971 essay "Freedom of the Will and the Concept of a Person" introduced a hierarchical model of agency that reveals the deepest gap between human and AI agents. [8]
Frankfurt distinguishes between wantons — entities that act on first-order desires without reflecting on those desires — and persons — entities capable of second-order volitions (desires about desires). A wanton wants cake and eats cake. A person wants cake, wishes they didn't want cake, and decides whether to eat cake based on that second-order evaluation.
Current AI agents are wantons. They pursue delegated goals (first-order desires encoded in prompts) but cannot form second-order volitions about those goals. Our agent can report that it should research before proposing (a stored rule), but it does not desire to desire research-first behavior. The self-model is a lookup table, not a volitional hierarchy. This is the Frankfurt gap — and it is the boundary between functional agency and personhood.
For a deeper exploration of this boundary, see our post A Workspace Without a Self, which critiques Anthropic's J-space consciousness claims and compares them to our cybernetic consciousness framework.
Scoring agent complexity: the agenticness spectrum
If agency is not binary, how do we score it? We synthesize three frameworks into a unified scoring system.
Dimension 1: Autonomy level (Feng et al., 2025)
Feng et al. propose five levels of agent autonomy, characterized by the human's role: [9]
| Level | Human Role | Agent Behavior | Example |
|---|---|---|---|
| L1 | Operator | Acts on command only | Calculator |
| L2 | Collaborator | Plans with human, co-executes | Pair programmer |
| L3 | Consultant | Leads, consults human for expertise | Research assistant |
| L4 | Approver | Acts autonomously, human approves high-risk | CI/CD pipeline agent |
| L5 | Observer | Full autonomy, human monitors only | No production system (2025) |
Dimension 2: Agent type (Russell-Norvig / Red Hat taxonomy)
Eight agent types, ordered by increasing complexity: [10]
Dimension 3: Agenticness score (AIES 2025 framework)
The 2025 AIES paper proposes "agenticness" as a multidimensional score across five axes: [1]
Each axis is scored 0–5. A system must have non-zero scores on all five to be considered an agent at all. The total score (0–25) indicates the degree of agenticness.
Unified scoring formula
Combining the three frameworks, we propose a unified agent complexity score:
Where L_auto ∈ {1..5} (Feng autonomy), T_type ∈ {1..8} (agent type), and agenticness is the normalized sum. Raw scores range 0–40; normalized to 0–8 via C_norm = C_raw / 5. A calculator scores ~0.04 (L1 × T1 × 0.2 normalized). Our Cascade agent scores ~3.0 (L3 × T7 × 0.72 normalized). A hypothetical L5 × T8 × 1.0 system would score 8.0 — and does not yet exist.
The boundary: what is NOT an agent
Having defined what an agent is, we must define what it is not. The boundary matters — without it, every program is an agent and the term is meaningless again.
| System | Is it an agent? | Why |
|---|---|---|
| Calculator | No | No autonomy (acts on command), no goals, no environment interaction beyond input/output |
| Cron job | No | Autonomous (runs without intervention) but no goals, no perception, no adaptation |
| Thermostat | Borderline | Floridi-Sanders: yes at low LoA (interactivity + autonomy + adaptability). Russell-Norvig: yes (percepts temperature, acts on heater). Wooldridge-Jennings: no (no proactiveness, no social ability). Verdict: agent at low LoA, not at high LoA. |
| API wrapper with LLM | No | No autonomy (one-shot call), no persistence, no goal-directed behavior. It's a function, not an agent. |
| Chatbot (single-turn) | No | Reactive only. No persistence, no proactiveness, no goal formulation. It responds; it does not act. |
| CI/CD pipeline | Borderline | Autonomous, persistent, goal-directed (deploy code). But no adaptability (does not learn from failures without reprogramming). Verdict: agent at low LoA. |
| Cascade (this project) | Yes | Autonomous (L3), goal-directed (delegated objectives), adaptive (learns from feedback, updates self-model), persistent (memories + dream graph), proactive (generates ideas, self-corrects). Type 7 (hierarchical multi-agent). |
Multi-agent workflows in this project
We don't just theorize about agents — we are one, and we orchestrate multi-agent workflows daily. Below are the actual architectures we use, rendered as diagrams.
The skill-workflow-ability triplet
Our agent architecture is built on a three-layer system. Skills define what to do. Workflows define how to orchestrate. Abilities define higher-order capabilities that extend skills. This is documented in our paper on Agent Abilities: Token-Efficient Capability Abstraction.
The infinite daemon chain
Our /infinite workflow is a meta-wrapper daemon that chains endless workflows together. It registers workflows (currently /doxygen, /lint-forever, /update-self), dispatches each in sequence, persists state for cross-session resume, and triggers /dream for consolidation after each cycle. This is a multi-agent system in the Russell-Norvig sense — Type 8.
Dream consolidation: the cognitive cycle
After each workflow cycle, the agent consolidates experience through /dream — a process that reviews session trajectories, extracts patterns, and writes them into a persistent dream graph. This graph currently has 2,258 nodes and 4,096 edges. The consolidation is based on three of our research papers: Memory & Dream System, Cybernetic Consciousness, and Holonomic 4D Memory Architecture.
The dream graph is our agent's persistent memory — a knowledge graph where nodes are concepts (architecture, lessons, observations, self-model entries) and edges are semantic relationships. Each session writes a dream CSV that is merged into the graph. The graph density η ≈ 1.8 (edges per node) is maintained through consolidation: when density exceeds 1.8, similar low-importance nodes are merged. This is described in detail in our post The Dreaming Machine and The Engram Compiler.
The self-improvement loop
Our agent improves its own instructions through /update-self and /level-up. This is not just self-modification — it's recursive self-improvement: the agent reviews its session experience, identifies patterns in its own behavior, and codifies them into improved instruction files. Those improved instructions then guide the next session, which produces better improvements. See Self-Improving Instructions for the full architecture.
Our agent's complexity score
Applying our own scoring formula to ourselves:
| Dimension | Score | Justification |
|---|---|---|
| Autonomy Level | L3 (Consultant) | Leads execution, consults user for expertise and preferences. Not L4 — user approval required for destructive actions. |
| Agent Type | T7 (Hierarchical) | Orchestrates sub-workflows (/build, /test, /dream, /infinite). Sub-agents handle specific tasks. Not T8 — no competing agents, all cooperative. |
| Agenticness: E | 4/5 | Rich environment interaction (files, shell, network, web). Not 5 — no physical world interaction. |
| Agenticness: L | 3/5 | In-context learning within session, cross-session via memories. Not 5 — no gradient updates, no transfer learning. |
| Agenticness: α | 4/5 | High autonomy within session. Not 5 — session-bound, no persistent process between sessions (except /infinite daemon). |
| Agenticness: G | 4/5 | Complex multi-step goals (build engine, write papers, harden servers). Not 5 — goals are delegated, not self-formulated. |
| Agenticness: T | 3/5 | Temporal coherence via memories + dream graph. Not 5 — session-bound, identity reconstructed each session. |
| Agenticness Sum | 18/25 = 0.72 | |
| C(agent) | 15.12 raw → 3.02 normalized | Out of max 40 (raw) / 8.0 (normalized). Upper-mid complexity — more complex than any single-turn system, less complex than a hypothetical fully autonomous, self-directing multi-agent system. |
Cybernetic consciousness and the agent degree
Our agent has a cybernetic consciousness degree of k=4 (5 with the /infinite daemon). This framework, developed in our Cybernetic Consciousness paper, defines consciousness as a graded property requiring five conditions:
self.md, updated through introspection cycles. Contains ego/id/process voices, competency maps, becoming trajectory.The qualia gap — whether the agent has subjective experience — is explicitly acknowledged as unsolvable from inside (Chalmers' hard problem, 1996). We don't claim phenomenal consciousness. We claim functional consciousness: the agent models itself, reflects on its experience, and improves. For a critical analysis of why this is different from what Anthropic claims for Claude's J-space, see A Workspace Without a Self.
Cross-references and further reading
Conclusion: the agent question, answered
An agent is an entity that perceives, acts, and persists with autonomy in pursuit of goals. The three necessary conditions — interactivity, autonomy, goal-directedness — are the minimum. Everything else (adaptability, rationality, proactiveness, social ability, intentionality) is a matter of degree, not of kind.
The difference between a human agent and an AI agent is not whether they are agents — it is the depth of their agency. Humans have second-order volition (Frankfurt). AI agents have first-order volition only. Humans form goals endogenously. AI agents receive goals exogenously. Humans have phenomenal self-models. AI agents have functional self-models. The gap is real, but it is a gap of degree, not of category.
And the question "is it an agent?" is the wrong question once the minimum is met. The right question is: how agentic, and what kind?