Contribution
The essay asks when a whole is more of an agent than its parts, not whether something is an agent. It answers in dynamical-systems terms. A whole is more agential when its internal organization makes the parts better at satisfying their own goals than they could be under the organization it had before. It then argues, from earlier models by the author and co-workers, that such organization can arise without design, from unsupervised learning by self-interested parts. It ends that “the agent creates itself”.
Key insight
Agency at one level is measured in the currency of the level below. A system’s energy function, its actual dynamics, can differ from the objective function given by the frustrated constraints among its parts. When the dynamics take the parts to configurations that satisfy more of those constraints than the obvious local equilibrium does, the whole is goal-directed with respect to its parts. Sometimes they do this by going uphill in the short term. “Even if the goals of the parts were defined arbitrarily, goal-directed behavior can be quantified objectively at the system level given the goals of the parts.”
Assumptions
- Mechanism granted. “We will suppose (without prejudice) that a system admits a mechanistic account in principle.” No vitalist line is drawn: “we avoid trying to identify any vitalist criterion that categorically distinguishes between agential and non-agential … systems in absolute terms.”
- A given decomposition into parts with goals. The norm for the whole is the parts’ norm. The author calls this “robbing Peter to pay Paul”, and accepts that “agency of the organism can be different or absent if we define the agency of the parts differently”. The result is “sensitive to different ways of defining its subagents”.
- Agency depends on what is observed. It “depends on what variables are observed and which are not”, with internal variables screened off from external ones.
- Behaviour as local descent on a landscape of constraints among parts, after Fields and Levin (2022). Frustration is the set of unresolved constraints.
- Low-order constraints for the learning mechanism. “[I]f the constraints are low order (e.g., pairwise), there will be a positive correlation between the quality of a solution and the size of its basin of attraction.” The Hebbian mechanism’s improvement relies on this.
Key results
The essay proves nothing new. Its results are definitions, a worked example, and reports of earlier models.
- The comparative 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).” Stated bottom-up: “a system of parts is agential if they are able to acquire organized relationships among themselves that exhibit goal-directed behavior greater than that which they exhibit as individuals.”
- 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.
- Goal-directedness as delayed gratification. For a system 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)”. A system “goes uphill in a specific direction that leads to lower points in the long term. It knows which way to go.” Stochastic escape does not count.
- The low end. A ball with momentum “can find better optima compared to a ball without momentum (in some environments)”. Minimal agency “still skates very close to no agency at all”. A system whose internal state makes its parts do better “is not at zero on the agency scale”.
- Under-determination (Box 1, Fig. 1). In a two-hidden-unit network for XOR, either hidden unit can compute any of the 8 functions sensitive to both inputs, and the network still classifies all four samples correctly. Two zero-error networks “only agree on half of the feature space”. So correct behaviour “tells us absolutely nothing about what any of its components computes”.
- Non-decomposable history (Box 2). Hidden-unit roles settle by a symmetry-breaking phase in which “the roles of h1 and h2 mutually define one another”. The resulting organization fixes generalization, and it “depends intimately on the history of the system, the whole system, and nothing but the system”.
- Learning to be an agent (reported from Watson et al. 2011a, c). A Hopfield-type network is shocked between attractors and its weights are changed by Hebbian learning. It enlarges the basins it visits most, which tend to be the better solutions, and so “finds the high-quality solutions more often” with respect to its original constraints. This needs no teacher or reward.
- Nonneural cases (reported).
- Utility-maximizing agents that adjust their relationships reach equilibria that “are actually better solutions to those original constraints” (Davies et al. 2011).
- Ecological networks do the same with “no population of ecosystems or group-level selection” (Power et al. 2015).
- Learned timing dependencies pick out subsets of agents to “work together”, rescaling the search to “a state space over emergent groups of variables” (Tissot et al. 2023, under submission).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | A whole’s agency can be compared with its parts’ by how far its organization resolves the parts’ mutual constraints | moderate | definition and argument, with the many-balls illustration |
| C2 | Goal-directed behaviour is going uphill in a specific direction to reach better optima, which requires dynamics that differ from the objective | moderate | argument from the ball-on-a-landscape example |
| C3 | Correct behaviour under-determines component organization, and the organization fixes generalization | strong | worked XOR example, Box 1 and Fig. 1 |
| C4 | A system’s organization can depend on its history as a whole, not decomposably on its parts’ histories | moderate | symmetry-breaking argument, Box 2 |
| C5 | Unsupervised, bottom-up learning among self-interested parts can produce organization that makes the whole more agential | moderate | reported from cited models (Watson et al. 2011a, c; Davies et al. 2011; Power et al. 2015), not shown here |
| C6 | Problem-solving competence of this kind is a fundamental feature of all organisms | weak | appeal to Levin’s experimental programme and Fields and Levin (2022) |
Concepts
- Frustration (stress). The state in which some parts’ constraints are resolved and others are “unresolved, or in tension”. A configuration’s quality is how many constraints it resolves, or its total energy or stress.
- Energy function against objective function. The function the system actually descends, including internal variables, against the function that scores solutions to the parts’ constraints.
- Agential with respect to its parts. A relative predicate. It holds when the whole’s organization makes the parts “do better at what they normally do” than a disorganized arrangement would.
- Self-modelling dynamical system. A dynamical system “with a behavior that is augmented by a model of its own past behavior”, acquired by Hebbian change to its connections.
- Conservative against innovative organization. The first holds on to good states already visited. The second also promotes good states never visited, which is generalization.
Connections
The essay appears in a Biological Theory thematic section on concepts of agency, beside Okasha (2023), whose “intelligent agent” sense it adopts. It builds on Fields and Levin (2022) for observer dependence and “patience”. Its own lineage is Watson’s evolutionary connectionism (Watson et al. 2011a–c, 2016; Watson and Szathmáry 2016; Kounios et al. 2016). It contrasts itself with Friston’s free energy: its information “makes its parts less frustrated … rather than … minimizes surprise”. It ties the account to the major transitions in individuality (Szathmáry and Maynard Smith 1995; Watson, Levin and Buckley 2022).
Bearing on the record
- For QUESTION-009: a comparative criterion, not a threshold. The LIT files it as “a candidate criterion for when coordinated parts constitute a further agent”. The essay does give a criterion. Coordination counts when it lets the parts resolve more of their mutual constraints than they would without it. But the criterion says when a whole is more agential than its parts, not when there is an additional agent. It declines any absolute line “regardless of whether the parts, or the system as a whole, are considered to pass some absolute threshold”. Read as an answer to QUESTION-009, it changes the question rather than answering it.
- The LIT summary says more than the essay. It says the notion “needs … distributed knowledge acquired by changing the organization of the parts”. The abstract says it “can involve” such knowledge. In the body, learned origin is an optional restriction: “Perhaps we might want to restrict a definition to systems that are able to acquire an internal organization”. The question of whether designed organization counts is left open. “Quantifiable” means only the comparison of total frustration at equilibria. The essay computes no quantity of its own.
- For QUESTION-032 and CLAIM-170. The account is explicitly level-relative and decomposition-relative. Agency 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”. This agrees with CLAIM-170’s factorization relative to a level. It does not supply QUESTION-032’s test for which factors are agents, because it takes the parts’ goals as given.
- For CLAIM-168, the circularity objection. Watson avoids presupposing a goal for the whole: “without presupposing that this is an aim for the system or, indeed, that the system is (already) an agent”. He does so by presupposing goals for the parts. The presupposition moves down a level and is not removed. That is the same shape as CLAIM-174’s reply, which takes candidates’ boundaries from the threshold criterion and not from agency.
- For CLAIM-174. Delayed gratification gives test (2) a behavioural mark: excursions uphill in the objective that reach lower minima, in a specific rather than a random direction. The Hebbian mechanism is an instance of test (3), a policy that changes with the consequences experienced.
- For CLAIM-010, the permissiveness objection. Watson bites the bullet at the low end: a ball with momentum is barely, but not zero, agential. Because the account is comparative, the bullet costs little. The thermostat (CASE-044) is not discussed. On this account it would need a decomposition into frustrated parts before the question arose.
- For CLAIM-172. The Tissot et al. result, reported second-hand, has learned relationships act on the timing of subagents’ actions. They pick out groups that “work together” and move the search to “a state space over emergent groups of variables”. That is close to CLAIM-172’s collective variable carried by mutual entrainment. The source was under submission and is not held here.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- The essay is a perspective piece. Its new material is conceptual. Every model-based result is reported from earlier papers, and none is re-derived or measured here.
- The measure needs a stipulated decomposition and stipulated goals for the parts. The author concedes these can be “defined arbitrarily”. The essay offers no way to choose among decompositions.
- The comparison is between two organizations of the same parts. It does not say what the baseline “disorganized” organization is in a real organism.
- The Hebbian result depends on low-order constraints and on the correlation between basin size and solution quality, which need not hold in general.
- The biological examples in the Discussion (regeneration, xenobot-like “skin cell clusters”) are cited to Levin’s work and not connected to the measure.
Open questions
- Is there a principled choice of parts, or does every decomposition give its own agency ranking? A result showing the ranking stable across reasonable decompositions would make the measure usable as a criterion.
- Does designed organization that resolves the parts’ frustrations make an agent? The essay raises this and leaves it open.
- Can the measure be computed for a real multicellular or social system, where the “disorganized” equilibrium is not observed?