aasb-ch05-agency
Chapter 5: Agency
Agency as a property of the loop
The Derivation
Attention answers "which gap do I address?" Agency answers a prior question: "do I generate the fetch, or does something else?"
A thermostat has a loop. It senses temperature, compares to setpoint, activates heating or cooling. But the thermostat doesn't choose its setpoint. It doesn't decide when to run. It doesn't evaluate whether heating is appropriate. The loop runs, but the thermostat doesn't own the loop.
Agency is the capacity to initiate the loop from within—to generate the FETCH rather than merely respond to external prompts.
| System | Loop Present | Fetch Origin | Agency |
|---|---|---|---|
| Thermostat | Yes | External (setpoint given) | None |
| Current LLM | Yes | External (prompt given) | None |
| Autonomous agent | Yes | Internal (gap self-detected) | Present |
The difference is not intelligence. The difference is who creates the hole.
Technical / Science
Active Inference
Karl Friston's Free Energy Principle describes organisms as minimizing "surprise"—the difference between predicted and actual sensory states.
But organisms don't just passively update predictions. They act on the world to make it match their predictions. You're cold; you don't just update your model to "I'm cold." You get a blanket.
Active inference is the loop closing through the world:
- PAUSE: Sense prediction error (gap between expected and actual)
- FETCH: Generate action that would reduce prediction error
- SPLICE: World changes, new sensory input arrives
- CONTINUE: Reduced prediction error, new state
The agent doesn't wait for information to arrive. The agent makes information arrive by acting.
Control Theory and Servomechanisms
A servomechanism is a device that uses error-sensing feedback to correct the action of a mechanism. The key insight: the controller produces the corrective action.
| Component | Control Theory | Agency Mapping |
|---|---|---|
| Reference | Desired state | What I want to be true |
| Sensor | Current state measurement | What I observe |
| Error | Reference − Sensor | The gap |
| Controller | Computes corrective action | The agent |
| Actuator | Executes action | The body/tools |
| Plant | System being controlled | The world |
The controller is where agency lives. Given an error signal, it generates the appropriate response. It's not told what to do—it computes what to do.
Goal-Directed Behavior
Goal-directedness is agency made operational. A system is goal-directed if:
- It represents desired outcomes
- It selects actions that move toward those outcomes
- It adjusts actions when outcomes diverge from goals
This requires the loop, but it requires more: the internal generation of the goal representation and the internal selection of the action.
A chess engine has goals (win the game) and selects actions (moves). But the goal was given externally. The agency, such as it is, operates only within the frame provided.
Full agency includes choosing the frame.
Autonomy Gradients
Systems exist on a gradient of autonomy:
| Level | Example | Fetch Origin |
|---|---|---|
| None | Rock | No fetch |
| Reactive | Thermostat | Setpoint external, sensing automatic |
| Deliberative | Traditional AI | Goals external, planning internal |
| Reflective | Meta-learning systems | Can modify own planning |
| Self-determining | ??? | Chooses own goals |
Current AI sits around level 2-3. The jump to level 4-5 is the jump to genuine agency.
Business / Practical
Decision-Making Under Uncertainty
Organizational agency is the capacity to decide and act despite incomplete information.
The agency gradient in organizations:
- Entry-level: Execute assigned tasks (fetch given)
- Manager: Choose how to accomplish goals (fetch partially internal)
- Executive: Set goals, allocate resources (fetch largely internal)
- Founder: Define what game to play (fetch origin)
Delegation is transferring agency—giving someone else the authority to generate their own fetches within a bounded domain.
Strategy as Choosing Which Loops to Close
Strategic decisions are meta-level agency:
- Which markets to enter (which gaps to care about)
- Which capabilities to build (which fetches to enable)
- Which competitors to ignore (which gaps to leave unfilled)
The strategist doesn't just operate the loop. The strategist designs which loops the organization will run.
Accountability: Who Owns the Loop?
When things go wrong, accountability asks: who had agency?
- If the loop was running on external instructions, accountability lies with whoever gave the instructions
- If the agent generated the fetch internally, the agent bears responsibility
This is why "I was just following orders" is a contested defense. It's a claim about fetch origin—"I didn't generate this; it was given to me."
Genuine accountability requires genuine agency. You can only be responsible for loops you could have chosen not to run.
Entrepreneurship as Agency Creation
The entrepreneur creates a new loop where none existed:
- Sees a gap no one is addressing
- Generates a fetch (starts a company)
- Splices resources together
- Continues until the gap is filled or resources exhaust
This is agency in economic form. The gap wasn't assigned. The fetch wasn't prompted. The entrepreneur initiated.
Theology / Philosophical
Free Will and Determinism
The free will debate maps to agency:
| Position | Fetch Origin | Implication |
|---|---|---|
| Hard determinism | All external (physics) | No genuine agency |
| Libertarian free will | Internal, uncaused | Agency from nowhere |
| Compatibilism | Internal, caused but "mine" | Agency within constraints |
The premises don't resolve this debate directly. But they clarify the question: Is there such a thing as internally generated fetch, or is "internal" an illusion—just more external causation we can't see?
Providence and Human Choice
Theological traditions grapple with divine foreknowledge and human agency. If God knows all outcomes, are human choices genuine?
One resolution: foreknowledge doesn't cause. God knows what you will freely choose, but the choice is still yours. The fetch is internally generated; divine knowledge doesn't make it external.
Another: divine will operates through human agency. Providence and freedom are not in competition—the loop runs, agency is real, and the outcome serves larger purposes.
"Faith Without Works Is Dead"
James 2:17 is a statement about agency. Faith (an internal state) without works (action) is incomplete. The loop must close.
Believing something matters requires doing something about it. Internal conviction generates external fetch. The loop runs through the world.
This isn't works-righteousness (agency earns salvation). It's the observation that genuine internal states produce observable external effects. If they don't, the internal state wasn't what you thought.
Co-Creation
Some theological frameworks describe humans as "co-creators" with God. This is agency language: genuine participation, not mere execution of predetermined script.
The loop runs. Human agency is real. The world changes because we act. And somehow this is compatible with divine sovereignty.
The how is mysterious. The what is clear: agency is taken seriously. Our fetches matter.
The Loop Property
Agency is the difference between:
| Loop runs | Gap source | |
|---|---|---|
| Reactive | External trigger | Environment |
| Agentic | Self-initiated | Internal generation |
But there's a deeper property: the delegation principle.
An agent alone cannot fully self-evaluate. The four premises force this:
- Embeddedness: no god's-eye view of yourself
- Incompleteness: your model of yourself is underdetermined
- Finitude: you can't store your full history
- Causality: you can't observe your future states
Therefore: self-evaluation has Gödelian limits. You cannot fully evaluate your own modifications using only internal reference.
The solution: External witness. Delegate evaluation to another agent with a different vantage point.
Love as control theory: Being willing to serve as someone else's sensor. To return accurate signal about their state when they can't see themselves. To close their loop when self-perception is compromised.
This isn't sentiment. It's the minimal viable solution to the self-reference problem.
See: delegation-principle for full development.
The Delegation Principle
From the premises, a crucial corollary: agents cannot fully evaluate their own modifications.
The Gödelian limit applies to self-reference. A sufficiently complex system can formulate questions about itself that it cannot answer from within. "Should I change this goal?" requires a reference point for evaluation—but if the goal is what you're evaluating, the reference point is the thing in question.
Resolution: Fetch externally. Ask another agent. Consult a human. Read a document you didn't write. Delegation is not weakness—it is the only way to break self-referential deadlock.
This is why relationship is computationally necessary (Part VIII of AYNL). An agent alone cannot fully self-evaluate. An agent in relationship can delegate evaluation to external witnesses.
Agency + isolation = eventual incoherence. Agency + relationship = sustainable self-modification.
Love is the willingness to serve as another's external evaluator.
Provenance
Document
- Status: 🔴 Unverified
South
slots:
- slug: delegation-principle
context:
- The Delegation Principle extends Ch5's agency content - shows why external witness
is structurally necessary, not optional