Contribution
A diagnosis of “category inflation” in talk of agentic AI. Kahl separates operational autonomy, which he accepts as engineering vocabulary, from “attributive agency”, agency for purposes of responsibility, authority and governance. He gives a structural criterion for the second, continuity pressure, positions it against five neighbouring accounts (Floridi and Sanders, Dennett, Dung’s dimensions, Bratman’s planning, Frankfurt’s endorsement), applies it to four classes of system and two borderline classes, and draws governance consequences: misattributed agency displaces responsibility, worst where affected persons cannot exit.
Key insight
Optimising well is not the same as owning the standard one optimises under. A system becomes a candidate for responsibility only when its standard itself can break down and the system, not its operator, has to settle what replaces it and live with the settlement. While failure is met by retraining, rollback, prompt changes or replacement, the evaluator is not at stake, and there is no agent there in this sense, however capable the system.
Assumptions
- A distinct attributive sense of agency. Thin operational agency-talk is left alone: “The present article accepts the usefulness of thin operational agency-talk in computer science and AI engineering. Its target is narrower: the migration from operational description to normative attribution in governance contexts.”
- Substrate neutrality. “nothing in the account requires biology, consciousness, personhood, or human-like intelligence.”
- Functional, not metaphysical, continuity. “The phrase ‘continuing evaluative standpoint’ is used functionally, not metaphysically. It does not require numerical identity, biological continuity, consciousness, or an enduring self.”
- Necessary only. “It offers a necessary condition, not a sufficient theory of agency, and makes no claims about free will, consciousness, moral responsibility, or desert.”
- Diagnostic method. “Its methodological posture is diagnostic rather than revisionary.” No formal model, no data.
Key results
- Capability does not fix where evaluative authority sits (§2). Table 1 pairs each functional property of agentic AI with what it does not show: planning not authorship of evaluative standards, tool use not ownership of objectives, memory not continuity pressure, multi-agent coordination not unified evaluative authority, autonomous action not responsibility-bearing agency. “Multi-agent coordination may distribute execution, but distribution of execution is not the same as unity of authorship.”
- Neighbouring accounts presuppose the evaluator (§3). Decision theory, planning and hierarchical endorsement “presuppose the existence of an agent whose attitudes are being organised”. Dung’s dimensions are in a different register: they say “what a system can do, whereas continuity pressure characterises what is at stake for the system in doing it.” “Capability and continuity are independent properties, not endpoints on a single scale.” Treating any of them as sufficient “reverses the order of explanation: it assumes what it ought to explain.”
- The criterion (§4). “Continuity pressure is not persistence over
time alone. It begins where an inherited evaluative frame no longer
settles how future operation should proceed.” The threshold “is
therefore not revision as such, but revision whose consequences must
be carried forward by a continuing evaluative standpoint.” Authorship
“marks the structural relation between frame-breakdown and attributable
frame-settlement.” Three exclusions: “Continuity pressure is absent
where revision is externally imposed, where rollback avoids carrying
forward the consequences of revision, or where the system merely
selects among revision paths already ranked by a fixed higher-order
criterion.”
- The doctor whose patient falls outside the protocol: “The issue is not correction within a rule, but whether the rule still governs the case.”
- Meta-learning, reward-model updating and constitutional constraints, if bounded by an antecedent rule, are “adaptation within an evaluative frame rather than authorship of that frame.”
- The test (§5). Three questions: does future operation depend on
preserving the same evaluative standpoint through revision; can failure
threaten that standpoint rather than trigger correction; is the revised
frame “settled under breakdown, rather than determined by an antecedent
objective, reward structure, policy, or deployment constraint?”
- Tool-using language-model agents fail by resetability: “The agent does not have to carry forward a revised standard as the basis for its future operation.”
- Reinforcement learners fail because revision stays bounded by “an antecedent admissibility structure”. AlphaZero’s novel strategy “emerged within an antecedent frame”. Preference-trained assistants are the “most important contemporary stress test”, and fail: “Internal representation of evaluative preference is not authorship of evaluation.” Reward hacking is diagnosed and corrected from outside.
- Multi-agent workflows: “unity of action is not unity of authorship.” Disagreement among planner, critic and tester “is resolved within an orchestration regime”.
- Institutional systems (triage, the Dutch childcare benefits scandal, COMPAS, the Allegheny screening tool) exercise “authority without authorship”.
- Borderline: a self-modifying skill-library agent (Voyager) is closer, because history becomes operationally relevant, but still bounded by a meta-objective and resettable. A continually learning clinical system is closer still, because rollback is costly, but “inertia is not authorship.”
- The group-agency reply (§5.3). Kahl grants that a workflow could be a List-and-Pettit group agent. What separates a court from an orchestration is that the court’s standards for its own integration rule “are themselves contestable through institutional process”, whereas a workflow’s rule “is revisable only through external intervention”. “The structural distinction is not whether the integration rule was originally ‘designed’, but whether revision of that rule under breakdown is internal or external to the candidate group agent.”
- The sceptic (§5). The worry that no artefact can meet the criterion, since makers can always reset it, is met by saying the criterion is “a structural specification of what would need to change”, “demanding by design, but not unfalsifiable.”
- Pseudo-agency (§6). Distributed inputs plus concentrated institutional closure make an output decision-settling, and agency language “migrates upward to the system as a whole”.
- Governance (§§7–8). Corrigibility, auditability, human authority over objectives and human accountability “presuppose that the system remains an object of governance rather than a bearer of evaluative authority.” “In deployed governance contexts, operational autonomy without agency is often a condition of governability rather than a defect.” Under non-exit, misattribution leaves “a gap in traceability between institutionally embedded evaluative standards and the actors responsible for specifying, relying on, and revising them.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Operational capacities (planning, tool use, memory, coordination, autonomy) do not establish who authors the evaluative standards | moderate | conceptual; Table 1 |
| C2 | Attributive agency requires continuity pressure: settlement of a revised evaluative frame under breakdown, carried forward by the same standpoint | moderate as a proposal | definitional, motivated by the doctor case and the presupposition argument |
| C3 | Revision ranked by a fixed higher-order criterion, imposed from outside, or escaped by rollback is not authorship | moderate | follows from C2 |
| C4 | Representative current systems ordinarily fail the test | moderate for typical deployments | case analysis, not measurement |
| C5 | What separates a group agent from an orchestration is whether revision of its integration rule under breakdown is internal, not whether the rule was designed | moderate | the court contrasted with the workflow |
| C6 | Governance frameworks presuppose that deployed systems are not agents in this sense | moderate | NIST, ISO/IEC 42001, audit literature |
| C7 | Misattributed agency displaces responsibility, most under non-exit | moderate | mechanism argued from cases |
Concepts
- Operational autonomy. Task execution with reduced need for prompting.
- Attributive agency. Agency for purposes of responsibility, authority and governance.
- Evaluative frame. The standards by which success, failure, relevance, risk or permissible action are assessed.
- Continuity pressure. The condition in which the inherited frame no longer settles future operation and a revision must be carried forward by a continuing standpoint.
- Authorship. The relation between frame breakdown and attributable frame settlement; not creation, intention or originality.
- Closure. The point at which an output is treated as decision-settling for institutional purposes.
- Pseudo-agency. Agency language attaching to a system because closure is organised around its output.
Connections
It positions itself against Floridi and Sanders (2004) and Floridi’s agency without intelligence, Dennett’s intentional stance, Dung (2025) and van Lier’s dimensional accounts, Butlin (2024) on reinforcement learning and agency, Bratman’s planning theory, Frankfurt’s hierarchy, List and Pettit’s group agency, and the responsibility-gap literature (Matthias; Santoni de Sio and Mecacci; Vallor and Vierkant). It cites no enactive or autopoietic work, and none of Barandiaran, Di Paolo or Moreno. Benthall and Salge (LIT-1031, not read) reach a neighbouring legal conclusion from the enactive side.
Bearing on the record
On QUESTION-092: Kahl’s answer is that the location of control matters, but not the location of the scheduler.
- Allocation of execution among a system’s own components is not what he counts. A multi-agent workflow allocates work internally, with “voting rules, ranking functions, confidence thresholds”, and he says “These structures coordinate execution; they do not generate authorship of the governing frame.” So internal authority allocation, as such, does not make the stronger difference.
- What does is whether the system revises its own integration rule and standards under breakdown. That is an internal/external distinction, and it does not reduce to behaviour: two workflows that behave alike differ if one’s rule is revisable only by its developers. The question’s summary expected Kahl to make “the scheduler’s location beside the point”. That was half right. The scheduler’s location is beside the point; the location of revision under breakdown is the point.
- It rejects origin as the criterion: “not whether the integration rule was originally ‘designed’”. That bears on CASE-081. On Kahl’s view the second stage counts if, and only if, the system settles revised standards under breakdown. A system that keeps maintaining itself on the learned policy after the reward stops has not yet shown that; it “merely selects among revision paths already ranked by a fixed higher-order criterion”, or does not revise at all.
On CLAIM-288: independent support, from a different notion of agency.
- Kahl argues from responsibility, not self-constitution, and never mentions enactivism. He reaches the same conclusion about systems whose objectives are fixed from outside: they “may optimise effectively while lacking authorship of those standards”. Self-modification bounded by “an antecedent meta-objective” does not change that. CLAIM-288’s second answer, that the modeller’s objectives are only initial conditions, is the case Kahl calls borderline, and his condition for crossing is stricter than the claim’s: revision settled under breakdown, not ranked by any antecedent criterion.
- CLAIM-285’s endorsement scheme computes authority by an aggregation rule no operator controls. On Kahl’s test, that is an orchestration regime, unless the population can revise the rule itself under breakdown.
On CLAIM-116. Kahl’s position is a version of it for evaluation: adaptation, meta-learning and self-modification are regulation within a frame, and do not establish an author of the frame.
On CLAIM-086. Kahl’s “authority without authorship” has the same shape: an institutional system’s output governs downstream action while the authority of its standards lies elsewhere. Kahl does not frame it as a claim about rules governing each other’s scope, so this is resemblance, not support.
On THEORY-071. Both deny that optimising an externally fixed function makes an agent. They are not the same account. THEORY-071’s normativity is set by a self-maintaining organisation; Kahl’s is authority over evaluative frames, with no metabolic or organisational grounding, and he expressly drops individuation by persistence (“not numerical identity, personhood, or selfhood”). A bacterium is an agent on THEORY-071 and plainly not on Kahl’s attributive sense. Different explananda, so no rivalry.
Limitations
- Conceptual throughout. The cases are described at the level of typical deployments; nothing is measured.
- The phrase “settled under breakdown, rather than determined by an antecedent objective” carries the whole criterion, and it is not made precise. Every revision process has some prior structure; how much antecedent ranking is too much is not said. The paper concedes the threshold is approached “along a small number of overlapping dimensions” that are “illustrative rather than canonical”.
- The sceptic’s worry is answered by assertion that the criterion is falsifiable, without a case that would satisfy it.
- The governance sections assume the conclusion of §5 and add institutional mechanism; they are not further evidence for the criterion.
- Humans are said to remain “bearers of judgement through correction”, but whether human standards are settled under breakdown rather than ranked by antecedent upbringing is not examined.
Open questions
- What would a system look like that settles a revised frame under breakdown without an antecedent ranking of revisions? Kahl gives no positive case, artificial or biological.
- Is the court’s internal contestability of its integration rule enough, or must the standards for that contest also be internal? The regress is not addressed.
- Does continuity pressure coincide with CLAIM-025’s capacity to transform the governing organization? It looks close, and a reading that set the two side by side could tell.