Source
Abel, Barreto, Bowling et al. (2025), “Agency Is Frame-Dependent”, LIT-969, read in full in NOTE-793: §1 (the four-part account), §2 (Claims 1–4 and the conclusion) and §3 (choosing a frame).
What was actually shown
An informal argument in four steps, each adapted from a cited result.
- Boundary (Claim 1). “Nearly all agents admit many plausible boundaries”, and quantities such as an RL agent’s optimal policy and Bellman error change with the boundary (Jiang 2019, Prop. 10).
- Source of action (Claim 2). “There exist cases with at least two plausible choices of causal variables” on which Kenton et al.’s agent-discovery method (LIT-909) gives opposite verdicts.
- Goals (Claim 3). Every input-output system can be described as if goal-directed, and the zero reward is consistent with every behaviour, so deciding that a goal is “meaningful” needs an extra principle (Occam, maximum entropy, Dennett’s usefulness).
- Adaptivity (Claim 4). Whether a fixed policy is adaptive depends on the reference class of behaviour changes counted (Zadeh 1963; Abel et al. 2023, Thm 3.1).
- Conclusion. Agency, as the conjunction of these conditions given a boundary, is “frame-dependent”. A frame is the four commitments together. No formal definition of a frame is given, and the paper says its arguments “are purely philosophical”.
Each step could have failed: a condition might have had a single natural reading that no plausible alternative contests. The paper offers one contested case or cited result per condition, and no more.
What this does not say
- Not that every attribution of agency is frame-relative. The premises are existential (“there exist cases”, “many”), and the conclusion is not. Systems on which all reasonable frames agree are not ruled out.
- Not that the frame is arbitrary in the sense of unprincipled. The paper asserts arbitrariness but argues only that a choice is needed. Its own §3 proposes predictive or explanatory power as a selection principle, which, if defensible, would fix the frame.
- Not that agency is unreal or merely conventional. Within a frame the conditions can be decided, and comparisons can be made.
- Not a refutation of self-individuating accounts. Barandiaran et al. (LIT-566), on which the four-part account is modelled, require a boundary and norms that the system’s own organisation produces “without an observer”. The paper’s examples are of observer-drawn boundaries and behaviour-inferred goals. It does not engage the organisational claim (THEORY-071).
- Not a test. It gives no criterion that admits some systems as agents and refuses others, and no verdict on the thermostat (CASE-044).
Connections
- QUESTION-032. Supports the question’s relativization to a level and
a target, and widens it. A frame also fixes the causal variables, a
principle for meaningful goals and a class of meaningful adaptations.
It is not an answer, and carries no
answersedge. - THEORY-232 (Watson, LIT-990). Watson’s comparison is relative to a
chosen decomposition and stipulated goals for the parts. Those are
frame components (1) and (3) here. This theory generalizes the
relativity Watson concedes for one measure to every account. It does not
touch comparison within a frame, and shows no instance for Watson’s
measure, so it neither satisfies nor refutes THEORY-232’s
promote_when. - CLAIM-174. Its tests 2 and 3 already concede an observer-relative prior and Kenton et al.’s frame-dependence. Claim 4 here adds that test 3 also depends on which changes count as adaptation. The frame-selection principle the paper floats, predictive power, is what test 2 uses.
- CLAIM-152 and CLAIM-170. The target-relativity of sufficiency-based boundaries and the level-relativity of agents are both instances of frame component (1).
- CLAIM-194. It agrees that coupling criteria cannot supply individuality. It differs on whether self-production makes a boundary observer-free, a point the paper leaves unaddressed.
- THEORY-071. A rival in aim, though the paper does not engage it directly; see “What this does not say”.