# kaal:position:2026-09-12-005

**Affirmed position.** The benchmark evidence Cao cites shows agents degrading over continuous runs as errors compound. Kaal's position is that this pattern is authority drift: every step passes its check while the whole run goes beyond what was authorized. Evaluation must cover the whole run, and authority tied to tasks must expire when the work stops.

**Status.** affirmed  **Published.** 2026-09-12

**Holds when.**

- Cao's paper is a position paper citing benchmark evidence, not an empirical test of these claims.
- The position is Kaal's own argument in the essay Prompts Don't Govern Agents. Reputation Does.
- The position does not validate Cao's figures, benchmark methodology, or reported benchmark results.
- The position is limited to continuous runs, aggregate authorization, and task-scoped authority.

**Current debate.** Zhenfeng Cao, Agentic Software: How AI Agents Are Restructuring the Software Paradigm (arXiv:2606.05608v2, 2026): https://arxiv.org/abs/2606.05608

**Extends.** kaal:claim:7314479-015: https://wulfkaal.github.io/claims/7314479-015

**Scholarly basis.** Wulf A. Kaal, Institutional Requirements for Sovereign Local Agent Runtimes (2026). SSRN: https://ssrn.com/abstract=7314479

**Source PDF sha256.** `debace24a155ae924a155b1fafe98856d98cf83689feff2f87a32f1c06171ce6`

**Topics.** consensus-and-security, governance-design, ai-and-agents

**Provenance.** Affirmed in kaalvis:2026-09-12:prompts-reputation-essay-six-v1 at https://wulfkaal.com/2026/09/12/prompts-dont-govern-agents-reputation-does/.

**Record type.** This is a dated commentary position that extends a scholarly corpus claim. It is not a verbatim claim extracted from the paper.

**Canonical form.** This markdown file is the canonical hashed representation of the position.