To Throw a Stone with Six Birds: On Agents and Agenthood

This paper proposes a type-correct framework within Six Birds Theory that operationalizes and empirically tests agency by distinguishing it from mere objecthood through four checkable components—ledger-gated feasibility, viability kernels, feasible empowerment, and packaging maps—demonstrated via reproducible experiments in a minimal ring-world that separate agenthood from agency without relying on goals or consciousness.

The Gist

The Big Idea: What is an "Agent"?

Imagine you are standing in a field. You pick up a stone and throw it. The stone flies, hits a target, and knocks it over.

• The Stone: The object you threw.
• The Throw: The action you took.
• The Result: The world changed because of your throw.

The paper asks a simple but deep question: What actually makes you an "agent" (someone who can act) rather than just a thing that gets pushed around by the wind?

Many people think being an agent means having a brain, goals, or feelings. This paper says: No. Being an agent is about physics and stability. It's about having a stable "package" that can make a reliable difference in the world, even when things get messy.

The author calls this framework "Six Birds Theory." Think of these "six birds" as six magical tools needed to build a machine that can truly "throw a stone."

---

The Six Birds (The Tools You Need)

To turn a chaotic pile of atoms into a stable "Agent" that can act, you need these six ingredients:

1. The Wrapper (Packaging): You need to draw a line around yourself. You need to know where "you" end and the "world" begins. Without this, you are just a cloud of gas, not a thing.
2. The Rules (Constraints): You can't do anything you want. You have a budget. Maybe you only have enough energy to move your arm once. These limits actually create the action. If you could do everything, you wouldn't be doing anything specific.
3. The Rhythm (Protocol/Holonomy): This is the fancy word for "Order Matters." If you walk North then East, you end up in a different spot than if you walk East then North. An agent needs to be able to use these sequences to create new possibilities that a simple robot couldn't.
4. The ID Card (Identity/Staging): You need a way to say, "This is me, at this specific time." It's like having a timestamp or a name tag that stays the same even if your clothes change.
5. The Safety Net (Viability/Closure): You need to be able to survive. If a little wind blows you apart, you aren't an agent. You need a "safety net" that keeps you in one piece.
6. The Wallet (Accounting/Transduction): You need a way to pay for your survival. If you get damaged (like a scratch on your car), you need a way to spend your "energy budget" to fix it. If you can't pay to fix yourself, you fall apart.

---

The Two Big Concepts: "Agenthood" vs. "Agency"

The paper makes a crucial distinction between two things that sound the same but are very different:

1. Agenthood (The Hardware)

• Analogy: Imagine a car parked in a garage.
• Meaning: Does the car exist? Is it intact? Does it have a battery and fuel?
• The Paper's View: "Agenthood" is just the ability to persist. It's the "Safety Net" and the "Wallet." If you can't fix yourself when you get damaged, you don't have "Agenthood." You are just a broken toy.

2. Agency (The Software)

• Analogy: Now, imagine you are driving that car. You turn the wheel left, and the car goes left. You turn it right, and it goes right.
• Meaning: Can you actually change the future?
• The Paper's View: "Agency" is the ability to make a difference. If you have a car but no gas (no budget), or if the steering wheel is broken (no control), you have "Agenthood" (the car exists) but no "Agency" (you can't go anywhere).

The Paper's Conclusion: You need the Hardware (Agenthood) before you can have the Software (Agency).

---

The Experiments: How They Proved It

The authors built a tiny, simple computer world (a "Ring World") to test these ideas. Here is what they found, using simple metaphors:

1. The "Repair" Experiment (The Wallet)

• Scenario: Imagine a robot walking on a path that is full of holes.
• Without Repair: The robot falls in a hole, gets damaged, and keeps walking until it breaks. It stops being a "thing."
• With Repair: The robot has a wallet. When it gets damaged, it spends money to fix itself.
• Result: The robot with the wallet stays stable. It becomes a "real object" that can be trusted. Conclusion: You need to pay to stay alive.

2. The "Fake Control" Trap (The Rules)

• Scenario: Imagine a clock on the wall that makes a ticking sound.
• The Trap: If you think the clock is you making the sound, you might think you have control. But you don't; the clock is just ticking on its own.
• Result: The paper shows that if you mistake a random schedule for your own choice, you think you have power, but you don't. Real agency requires that you are the one pulling the lever, not just watching the world happen.

3. The "Order Matters" Experiment (The Rhythm)

• Scenario: Imagine a dance.
• Result: If you do Step A then Step B, you end up in a different place than if you do Step B then Step A.
• Finding: The paper showed that agents can use these "sequences" to reach places they couldn't reach with just a single step. It's like a chess player who plans three moves ahead, not just one.

4. The "Learning" Experiment (The Skill)

• Scenario: Imagine a clumsy person throwing a stone vs. a pro athlete.
• Result: The pro athlete (high skill) can throw the stone exactly where they want. The clumsy person (low skill) misses.
• Finding: "Learning" isn't just remembering facts; it's rewriting the physics of your own body to make your actions more reliable. The more you learn, the more "power" (agency) you have.

---

The Takeaway: "An Agent is a Theory Object"

This sounds complicated, but it's actually simple:

• The Theory: The set of rules and physics that define a specific world (like the rules of a video game).
• The Object: A character in that game that is stable, has a budget, and can make choices.

The paper argues that an agent isn't a magical soul. It is just a stable package that exists inside a set of rules.

• If the package falls apart (no repair), it's not an agent.
• If the package can't make choices (no budget), it's not an agent.
• If the package can't change the world (no difference-making), it's not an agent.

In short: To "throw a stone," you first need to be a solid thing that doesn't fall apart, and you need to have enough energy to actually throw it. If you have those two things, you are an agent. Everything else (consciousness, feelings, goals) is just extra decoration on top of this solid foundation.

Technical Summary

1. Problem Statement

The paper addresses the difficulty in empirically distinguishing agency (the capacity to make a difference in the world) from mere persistence (the ability to exist as an object). In current discourse, these concepts are often conflated, leading to:

• False Positives: Systems that appear active but lack control (e.g., a single-action system or a system following an exogenous schedule).
• Modeling Artifacts: Claims of agency that arise from misidentifying external noise or schedules as internal choices.
• Lack of Formalism: A lack of "type-correct" definitions for agents that do not rely on unobservable internal states like goals, consciousness, or utilities.

The core question is: What structural conditions must be met for a subsystem to be considered an "agent" capable of "throwing a stone" (making a reliable, counterfactual difference in the outside world)?

2. Theoretical Framework: Six Birds Theory (SBT)

The paper grounds its analysis in Six Birds Theory, which posits that macroscopic objects and structures are not primitives but are induced by layers of closure. The theory relies on six emergence primitives:

1. P1 (Operator-write): Rewriting effective transition laws (learning).
2. P2 (Feasible-set-write): Constraints defining admissible actions.
3. P3 (Protocol holonomy): Non-commutative action sequences (order matters).
4. P4 (Quantized identity/staging): Discrete tokens and time horizons.
5. P5 (Order-theoretic closure): Packaging stability (persistence).
6. P6 (Resource transduction/accounting): Budgeting and maintenance.

The Core Thesis: An agent is not a layer itself, but a "theory object"—a maintained package inside an induced theory layer.

• Agenthood: The enablement claim (the existence of a stable layer with a ledger-gated interface).
• Agency: The causal claim (the ability of that interface to induce different outside futures).

3. Methodology

The authors instantiate these abstract concepts in a minimal, exact finite-state "ring-world" environment. This substrate allows for mathematically rigorous, audit-able verification without the noise of continuous approximations.

Operational Definitions & Metrics

The paper defines three orthogonal metrics to test the thesis:

1. Viability Kernel ($K$): A robust, controlled-invariant safe set computed as the greatest fixed point of a monotone operator. A state is in $K$ if there exists a feasible action whose entire successor support remains safe. This measures persistence (Agenthood).
2. Feasible Empowerment: The channel capacity (mutual information) between feasible action sequences and an outside output variable, restricted by a ledger-gated budget. This measures difference-making (Agency).
3. Packaging Endomap & Idempotence Defect: An empirical map $E: X \to X$ on macro-labels (hiding micro-degrees of freedom). The idempotence defect ($Def(E)$) measures how often $E(E(x)) \neq E(x)$. A defect near 0 indicates the macro-labels behave like stable objects.

Experimental Design

The authors use a "toggle suite" to enable/disable specific SBT primitives (e.g., turning on/off repair, protocol holonomy, or learning) and measure the impact on the three metrics. They also include Null Regimes to guard against false positives:

• Null A: A system with only one action (no choice).
• Null B: A system where an exogenous schedule is mis-modeled as a controllable action.

4. Key Results

The experiments yield four distinct, checkable separations:

A. Maintenance is Required for Objecthood (Exhibit 4)

• Finding: When a "repair" action (P6) is disabled, the idempotence defect of the macro-lens remains maximal (1.0) at the staging horizon ($\tau=2$). When repair is enabled, the defect collapses to 0.0.
• Implication: Objecthood is not inherent in the lens; it requires active maintenance (spending budget to restore coherence) to stabilize the induced macro-description.

B. Null Regimes Prevent False Positives (Exhibit 5)

• Finding:
  • Single-action systems yield 0 bits of empowerment at all horizons.
  • Exogenous schedules, if correctly modeled as state/noise, yield 0 bits. If incorrectly modeled as actions, they falsely yield 1 bit.
• Implication: Agency requires a controllable interface variable; motion alone or mis-modeled external structure does not constitute agency.

C. Protocol Holonomy Enables Horizon-Dependent Control (Exhibit 6)

• Finding: At horizon $H=1$, empowerment is identical with or without protocol holonomy (P3). For $H \ge 2$, enabling non-commutative action sequences (order matters) significantly increases empowerment (e.g., from ~1.08 to ~1.69 bits).
• Implication: Agency can emerge from the geometry of action composition (order of operations) even if single-step capabilities are unchanged.

D. Learning as "Causal Thickening" (Exhibit 9)

• Finding: Increasing a discrete skill variable $\theta$ (P1, operator rewriting) monotonically increases median feasible empowerment (from 0.73 to 1.34 bits).
• Implication: Learning is not just storing information; it is rewriting the transition operator to make interface interventions more reliable, thereby increasing the capacity of the action channel.

E. The Phase Diagram of Agency (Exhibit 8)

• Finding: As internal noise increases, the viability kernel ($|K|$) and empowerment collapse to zero unless the cost of maintenance (repair) is low enough to be feasible.
• Implication: Agency is a fragile state that exists only within a specific region of the (noise, cost) parameter space where maintenance is affordable.

5. Key Contributions

1. Formal Definition of an Agent: Proposes the definition of an agent as a "theory object"—a maintained package with a ledger-gated interface inside an induced macro-physics layer.
2. Operational Metrics: Connects the abstract "Six Birds" primitives to concrete, computable metrics:
   • Viability Kernel size ($|K|$) for persistence.
   • Feasible Empowerment for causal leverage.
   • Idempotence Defect for object stability.
3. Guardrail Methodology: Introduces rigorous null regimes (single-action, schedule trap) to prevent the conflation of dynamics with agency.
4. Reproducibility & Formal Verification:
   • Provides a Lean 4 formalization proving that the viability kernel computation is indeed a greatest fixed point.
   • Publishes hashed configuration artifacts and deterministic scripts to ensure all quantitative claims are auditable and reproducible.

6. Significance

This paper shifts the study of agency from philosophical speculation or goal-based modeling to structural and operational verification.

• Decoupling: It successfully separates "being an object" (persistence/viability) from "being an agent" (difference-making/control), showing that one can exist without the other.
• Mechanistic Insight: It demonstrates that agency is not a magical property but an emergent phenomenon requiring specific structural ingredients: maintenance (to exist), constraints (to define actions), protocol (to enable complex control), and learning (to enhance reliability).
• Auditability: By using finite-state exact kernels and formal proofs, the paper offers a blueprint for "audited agency," where claims of autonomy can be verified mathematically rather than inferred from behavior alone.

In conclusion, the paper argues that to "throw a stone" (make a difference), a system must first be a stable, maintained package (Agenthood) and then possess a controllable interface that can reliably alter future states (Agency). Both conditions are necessary and can be rigorously tested using the proposed metrics.