nucleation

LIT-1033: How continuity distinguishes autonomy from agency in agentic AI

2 min read

Properties4
StatusActive — read in full 2026-10-13 (NOTE-723) from the publisher's open-access PDF (link.springer.com/content/pdf/10.1007/s44163-026-01675-5.pdf, CC BY-NC-ND), article 6:1187. Worth reading for a sharp separation of operational autonomy from agency that bears on responsibility, and for a test, continuity pressure, that turns on whether a system settles and carries forward revised evaluative standards when its inherited ones break down (THEORY-219). It supports CLAIM-288 from a responsibility argument rather than an enactive one, and gives QUESTION-092 a partial answer: internal allocation of authority does not count, internal revision of standards under breakdown does.
Tagsagency, self-governance, anthology-candidate
Cited as “paper” byNOTE-723 — Kahl — How continuity distinguishes autonomy from agency in agentic AI
Cited as “source” byTHEORY-1165 — Kahl proposes that agency in the sense that bears on responsibility needs a standpoint that settles and carries forward revised evaluative standards when its inherited ones break down, which is absent where revision is imposed from outside, escaped by rollback or chosen by a fixed higher-order rule; that allocating authority among a system's own components coordinates execution without authoring its governing frame; and that planning, autonomy, memory and coordination do not supply that standpoint, THEORY-219 — Agency in the sense that bears on responsibility needs a standpoint that settles and carries forward revised evaluative standards when its inherited ones break down, and planning, autonomy, memory and coordination do not supply one

Graph View

Kahl, P. (2026). How continuity distinguishes autonomy from agency in agentic AI. Discover Artificial Intelligence 6. DOI 10.1007/s44163-026-01675-5

Key takeaways

Read in full (NOTE-723).

  • Two senses of agency. Thin operational agency-talk is accepted as engineering vocabulary; the target is its migration into claims about responsibility, authority and governance.
  • Capability does not locate evaluative authority. Planning, tool use, memory, multi-agent coordination and autonomous action each leave open who authors the standards: “distribution of execution is not the same as unity of authorship.”
  • Continuity pressure. It “begins where an inherited evaluative frame no longer settles how future operation should proceed”, and it is absent where revision is imposed from outside, escaped by rollback, or chosen among paths “already ranked by a fixed higher-order criterion”.
  • The cases. Tool-using agents are resettable; reinforcement learners, preference-trained assistants included, revise under an antecedent admissibility structure; workflows coordinate without authorship; institutional systems have “authority without authorship”. Skill-library agents and continually learning clinical systems come closer and fail.
  • Group agents. What separates a court from an orchestration is whether revision of the integration rule under breakdown is internal, not whether the rule was designed.
  • Governance. Current frameworks treat deployed systems as non-agents, and misattributed agency displaces responsibility, most where affected people cannot exit.

Why this is here

A11 asked whether a system is self-governing ‘merely because authority allocation is internal, or must it retain responsibility for the standards by which its own decisions are evaluated’. Kahl answers the second: voting rules, ranking functions and the like “coordinate execution; they do not generate authorship of the governing frame”; only revision of the system’s own standards under breakdown does (THEORY-219). That supports CLAIM-288, by an argument from responsibility that cites no enactive work, and bears on QUESTION-092 and CASE-081, where it sides with present organisation against origin. The filing’s summary was right in outline. QUESTION-092 expected the paper to make “the scheduler’s location beside the point”; it does, but it makes the location of revision the point. The paper is a necessary condition stated without a positive case, and its key phrase, settled under breakdown rather than by an antecedent rule, is not made precise. Flagged anthology-candidate because the anthology’s agents-and-environments topic might hold a paper about AI agents; it is filed here for the question of evaluative authority (ADR-005).

Backlinks