nucleation

CLAIM-174: A candidate factor is an agent at a level when it is one system there, when a goal-directed description beats a device description by more than the Occam cost, and when its policy would adapt if its actions' consequences changed; it is a willed agent when, further, a will-organization over several regimes describes it better than none, and what passes that is the jurisdiction of a will-organization

7 min read

Properties16
StatusProposed
Version7
Rolethesis
Warrantinterpretive
Tagsagency, individuation, model-comparison
Linedistributed-agency
AnswersQUESTION-032 — Relative to a level of description and a target, what makes a factor of a system an agent, and what further makes it a willed agent?
Rests onCLAIM-157 — Relative to a stated target, a combined form is one system while its joint dynamics are nearly decomposable with the form as a block whose gap beats the Occam factor, and it lasts as long as that block persists; the unit is the gap-isolated block, not any member, CLAIM-154 — A candidate structure is worth naming as a unit when its effect-to-noise ratio outruns the Occam cost of keeping it: reify when d'² exceeds the log prior-to-posterior variance ratio, where the model with the candidate stays regular at the parameters that switch it off
GroundsLIT-911 — Agents and Devices: A Relative Definition of Agency, LIT-909 — Discovering Agents, LIT-398 — True Believers: The Intentional Strategy and Why It Works
SupportsCLAIM-367 — Whether a training-time regulatory condition belongs to a network depends on which unit is individuated for the training dynamics and whose jurisdiction governs the updates; in conventional training most of that governance falls to the optimizer–model–data system, so a condition of training is not thereby a condition of the network
Objects toCLAIM-168 — Identifying agents by their will-organization is circular: will-organization is defined for an agent already given, and the agent-versus-device comparison presupposes an identified system, so neither can produce the factorization it is meant to find
Objected byCLAIM-194 — A gap or coupling criterion cannot tell a boundary a system produces from one imposed on it: a membrane the cell makes and a beam joining two pendulums pass alike, so the block test cannot by itself supply the individuality minimal agency needs, CLAIM-393 — Ordinary and legal attributions of agency to an artifact cannot serve as the independent marks that would defeat the record's agency test, because they turn on whether its mechanism is understood: a known mechanism invites "This is all just programmed!", an inscrutable one invites fear, whatever the description lengths
ComplementsCLAIM-170 — The agents in a system are a factorization of it relative to a level of description and a target; agents in general are identified by adaptation, and willed agents are individuated by their will-organization: a willed agent is the jurisdiction of a will-organization, CLAIM-356 — An objective that originates outside a learning system can become a goal of the training organization, through its causal role in selecting, stabilizing and reproducing the learned structure it rewards; exogenous origin does not disqualify a goal, but it does not show that the network governs or endorses the objective, and what is selected is performance on the updating distribution, not generalization
Corroborated byTHEORY-092 — Behavioural capacity does not fix intelligence: a system with the capacity to respond sensibly to every possible input can lack it, so whether behaviour is intelligent, or the mark of any mental property, depends partly on how it is produced
UsesTERM-047 — jurisdiction of a will-organization, TERM-023 — will-organization, TERM-049 — combined form, TERM-044 — agent (at a level of description), TERM-048 — willed agent, TERM-046 — notation for the threshold, drive and unity criteria
Illustrated byCASE-044 — The thermostat: Dennett's knife edge between agent and device

Graph View

The claim

Fix a level of description and a target. Take a candidate factor of the system. It is an agent at that level when:

  1. It is one system there. It clears the threshold criterion (CLAIM-157). Its boundary comes from coupling, not from agency.
  2. The intentional description wins. A model in which the factor pursues goals predicts its behaviour better than a device model (an input-output mapping), by more than the Occam cost (TERM-046) of the richer model. This is Orseau et al.’s agent-versus-device comparison (LIT-911), priced in the record’s terms (CLAIM-154). Its ancestor is Dennett’s intentional strategy (LIT-398), which counts a system a believer when the intentional description predicts it better than what is antecedently known, and which Orseau et al. turn into a computation.
  3. Its governance would adapt. Under an intervention on how its actions affect the world, the factor’s policy, or what governs its regimes, would change. This is Kenton et al.’s criterion (LIT-909): agents are “systems that would adapt their policy if their actions influenced the world in a different way”.

What passes 1–3 is an agent (TERM-044). It is a willed agent (TERM-048) when it also passes:

  1. Will. A model with a will-organization, a maintained governance over several of its regimes deciding which acts, beats the agent model without one, by more than its Occam cost.

What passes 4 is the jurisdiction of a will-organization (TERM-047). That is the owner’s thesis made testable (CLAIM-170, CLAIM-155).

How it handles the cases

  • The thermostat (CASE-044).
    • Alone it fails (3). “An existing thermostat will not adapt to this change, and is therefore not an agent by our account.” On (2) it is on Orseau’s knife edge.
    • With its designers it passes. Kenton et al.: “the thermostat with its creation process is an agent under our definition”. On the jurisdiction reading, the thermostat’s one regime lies inside the designers’ will-organization. The agent is the jurisdiction, and its boundary includes the process that governs the device. The two accounts agree, and this is the strongest outside support for individuating agents by jurisdiction.
  • A trained policy against its training process. Kenton et al. call the training process, not the frozen policy, the agent: “simple forms of agents often derive much of their goal-directedness from the process that creates them”. This is the same structure.
  • The microbiome can pass (1) as part of the combined form (TERM-049). It fails to be the person’s agent, because nothing in the person’s will-organization governs it (TERM-047). Whether the community is an agent at its own level is a separate application of the test.
  • The dancing couple. At the level of the room, during the dance, it may pass all three: it is one block, the dance’s conventions govern both partners’ regimes, and the pair re-organizes when the floor or the music changes. At the level of the pair, the lead and the follow each pass. Kenton et al. find the same kind of plurality: an actor-critic is two agents, and the one-agent graph “is also an accurate description of the same underlying system, though with less detail”.

The two objections it answers

  • Circularity (CLAIM-168). The test runs on candidate factors whose boundaries come from (1). Only after (2) and (3) is a will-organization attributed, and then TERM-023’s agent index is satisfied by the admitted factor. The circle becomes a fixed-point condition. Several factorizations can satisfy it at once, and on this claim that is the level-relativity the owner predicted, not a defect. Whether that is so is the objection’s open part.
  • Permissiveness (CLAIM-010). The will paper answered it by supposing the agent (CLAIM-080). Here a control system without governance over its own regimes, or one whose rule would not adapt, is refused by (2) or (3). The bare thermostat is refused, and CLAIM-080 is not needed. CLAIM-080 stays Active for the will paper’s own scope.

What it does not say

  • That the test has been run on any real candidate. It needs interventional data on mechanisms, which Kenton et al. assume is available. For people and firms such data is hard to get.
  • That it is prior-free. Test (2) inherits Orseau et al.’s observer-relative prior, and test (3) inherits Kenton et al.’s frame-dependence: their labels “are only meaningful relative to the choice of which variables are included in the model”. Both are the level and target relativity this record already states (CLAIM-152). Test (3) also depends on what counts as adaptation (Abel et al., LIT-969, Claim 4; THEORY-249): Kenton et al.’s intervention on how its actions affect the world is one choice of meaningful adaptation, not the only one.
  • That an agent makes its own boundary. Test 1 reads the boundary off coupling, and a boundary kept from outside passes it as one the factor produces does (CLAIM-194). Tests 2 and 3 do not exclude such a factor in general: a policy can adapt inside a boundary someone else keeps. An agent here is TERM-044’s, at a level of description. Minimal agency in Barandiaran, Di Paolo and Rohde’s sense (THEORY-1119; the rival TERM-063) adds individuality the system produces itself, and this test does not claim it. Where the governor of an imposed boundary is found, the jurisdiction reading puts it inside the agent (note of 2026-10-21).
  • That what people or courts grant an artifact is an independent mark. Such grants turn on whether its mechanism is known (CLAIM-393). So they meet the defeated_if only where knowing the mechanism makes a device description shorter, which is test 2’s own comparison.

Note of 2026-10-12: evidence now linked

THEORY-092 (Block) now grounds this claim’s third test, weakly. Systems alike in all actual and counterfactual behaviour can differ through the processing, so a behavioural test alone admits a lookup table. Block’s subject is intelligence, not agency.

Note of 2026-10-21: test 3 and an imposed governance

A27 of the exchange of 2026-10-20 found “an important gap between conditions 3 and 4” and proposed to separate behavioural adaptation from governance adaptation. Test 3 already covers both (“the factor’s policy, or what governs its regimes, would change”), and A27’s thermostat repair is this claim’s own case with its designers. A27’s real point concerns the exchange’s governance-response profile (TERM-067), not this test. A description that preserves how a system would respond “if its governance mechanism were externally altered” does not show “that the system can alter the mechanism itself” (“An ordinary machine can respond dramatically to a technician rewiring it”). Test 3 is a test of the second, and the profile cannot stand in for it; CLAIM-407 says the same in its “What it does not say”. Where a system’s authority allocation is fixed from outside, this claim’s answer is the jurisdiction (TERM-047): the governor lies inside the willed agent’s boundary, and the system alone is not that agent. A27 did not see this. It also proposed “epistemically adaptive self-governance” as a further property, and rightly not as a condition: it “should not silently redefine willed agent”.

Note of 2026-10-29: the narrowed rule

CLAIM-154 now holds only where the candidate’s model stays regular at its switch-off. Where it is singular the rule overcharges (CLAIM-188), so a candidate it rejects may still earn a name.

Note of 2026-10-29: what holds it back

Its first condition is CLAIM-157, which is Proposed, and the necessary half of that criterion is unanswered (CLAIM-157’s note of 2026-10-29). With CLAIM-194 and CLAIM-393 written in, nothing else on file holds it back.

Backlinks