Source
Watson (2023), LIT-990, Biological Theory 19:22–36, read in full in NOTE-747: the criterion and measure (main text), the delayed-gratification argument, Boxes 1 and 2, and the Discussion.
What was actually shown
An argued definition with illustrations. The essay proves nothing new.
The criterion. “If a system at a higher level of organization causes the parts it contains to do what they ‘normally do’ but more, as if the complex constraints between them (that would have otherwise prevented this) were alleviated, then that system is agential (with respect to its parts).”
The measure. Agency is evaluated “by comparing the total frustration or energy of the system at local equilibria given some original organization of the parts to that obtained with some new organization of the parts.” No formula or scale is given beyond this comparison.
The behavioural mark. To beat a local optimum, “the energy function (describing the dynamics of the system) must be different from the objective function (describing the quality of solutions to the problem)”. The system “goes uphill in a specific direction that leads to lower points in the long term”. Stochastic escape does not count. A ball with momentum, which finds better optima than one without in some landscapes, is just above zero.
The relativity, stated by the author. The whole’s norm is borrowed from the parts’ goals (“robbing Peter to pay Paul”). The result is “sensitive to different ways of defining its subagents”, and the organism’s agency “can be different or absent if we define the agency of the parts differently”. Given the parts’ goals, though, “goal-directed behavior can be quantified objectively at the system level”.
One worked example (Box 1). In a two-hidden-unit XOR network either hidden unit can compute any of 8 functions while the network still classifies all four samples, so correct behaviour does not fix component organization, and the organization fixes generalization. This is the essay’s only result shown in full, and it supports the claim that the whole’s organization is a further fact about it, not the criterion itself.
What this does not say
- It does not say when there is an additional agent. The criterion is comparative. It declines any line “regardless of whether the parts, or the system as a whole, are considered to pass some absolute threshold”. As an answer to QUESTION-009 it changes the question.
- It is not observer-free. The comparison needs a chosen decomposition and chosen goals for the parts, which the author concedes can be “defined arbitrarily”. No way of choosing among decompositions is offered, and different choices may rank systems differently.
- It does not say that agency requires learned organization. Learned origin is offered as an optional restriction (“Perhaps we might want to restrict a definition to systems that are able to acquire an internal organization”). Whether designed organization counts is left open.
- It does not show that bottom-up learning produces agency. The Hebbian self-modelling network (Watson et al. 2011a, c), the self-interested agents (Davies et al. 2011) and the ecological networks (Power et al. 2015) are reported from earlier papers, not re-derived. The Hebbian improvement depends on low-order constraints and on a correlation between basin size and solution quality, which need not hold in general.
- It does not compute the measure for any real organism. The baseline “disorganized” equilibrium is not observed in a real multicellular or social system, and the biological examples in the Discussion are not connected to the measure.
Connections
No relations are declared: it answers a comparative version of QUESTION-009, not the question.
- QUESTION-009 and QUESTION-032. It gives a criterion for when coordination makes a whole more agential than its parts. It does not supply QUESTION-032’s test of which factors are agents, because it takes the parts’ goals as given.
- CLAIM-170. Agrees with its relativity to a level: agency here is relative to a decomposition.
- CLAIM-168 (circularity). The essay avoids presupposing a goal for the whole by presupposing goals for the parts. The presupposition moves down a level; it is not removed.
- CLAIM-174. Delayed gratification gives the goal-directedness test a behavioural mark: excursions uphill in the objective, in a specific direction, that reach lower minima.
- CLAIM-010 (permissiveness). The essay bites the bullet at the low end (a ball with momentum is barely, not zero, agential), and because the account is comparative the bullet costs little.
- CLAIM-172. The reported timing result (Tissot et al., then under submission, now held as LIT-989 and read in NOTE-773) has learned relationships pick out groups that “work together”, close to a collective variable carried by mutual entrainment. The learned relationship there is phase timing, not coupling weight, and the groups it finds are modules the couplings already define (THEORY-265).
- THEORY-071 (minimal agency). A different kind of account. Barandiaran et al. set absolute conditions, including individuality with no observer doing the distinguishing and norms the system’s own organization sets. Watson’s account sets no threshold and borrows its norm from a stipulated decomposition. They answer different questions and are not declared rivals.
- THEORY-249 (Abel et al., LIT-969, read in NOTE-793).
It generalizes this account’s frame-dependence. Watson’s comparison is
relative to a chosen decomposition and stipulated goals for the parts,
which are two of Abel et al.’s frame components; Watson concedes that
relativity for one measure, and Abel et al. argue that every account of
agency has it. Watson’s measure is comparative and stays meaningful
within a frame, which Abel et al. do not dispute. They predict no
ranking stable across decompositions in general, but show no instance
for this measure, so they do not settle its
promote_when.
Note of 2026-10-18: frustration as the record’s primitive, and what this reading’s measure aggregates
The exchange of 2026-10-17 made frustration, not distress, the primitive of the record’s account of regulatory difficulty (CLAIM-368, TERM-060), and A34 called this reading “the most important additional connection”. Three things bear on it.
- The measure is an aggregate. Watson compares “the total frustration or energy of the system”. A34 §2 concluded: “A higher-level organization should not be credited with resolving constituent frustration merely because its aggregate objective improves”, and offered it as following “directly from combining Watson’s proposal with Nucleation’s existing distinction between the good of a whole and the good of its constituents”. It does not follow from Watson. It is a criticism of his measure as read here: a lower total can come from sacrificing one part’s goal, which is the “robbing Peter to pay Paul” the essay itself names. CLAIM-347 gives the record’s refinement: compare which sets of the parts’ demands are jointly satisfiable, by inclusion, and leave two organizations unranked when each relieves a conflict the other does not.
- The decomposition-relativity is half of A38’s principle. A38 §2: “The attribution of frustration relief must be indexed to an individuation and an intertemporal correspondence of demands”, which A38 called “a defensible new formal principle”. The first half is this reading’s, already in its title (“relative to a chosen decomposition and the parts’ goals”). The second, a correspondence of demands across time when the parts change, is new, and is CLAIM-361’s.
- Delayed gratification is the precedent for resolve. The behavioural mark here, an energy function different from the objective that “goes uphill in a specific direction that leads to lower points in the long term”, is what A35 §4 meant by a system that may “deliberately worsen some objective now”, and what A37 and A39 developed as resolve. None of A34–A40 cited this reading for it.
Note of 2026-10-28: what holds it back
Kept Proposed by the sixth status review. NOTE-747 read the essay in full; it computes the measure for no system,
and Box 1 is its only result shown in full. Nothing in the record computes
the frustration comparison on one system under two decompositions, which
is what promote_when asks. THEORY-249 predicts instability in general but
shows no instance.
CLAIM-368 now cites this reading as corroboration, not as a ground.