The Institutional Boundary in Agentic AI

Agentic artificial intelligence is moving from content generation toward tool use, transaction initiation, system modification, and other forms of consequential action. This working paper tests a bounded hypothesis: several communities may be encountering the same institutional boundary—the transition at which governed computational activity becomes a basis for institutional consequence and downstream reliance.

The prospective evidence consists of three arm’s-length 2026 artifacts: the World Economic Forum and Kearney’s AI-first operating-system report, the Google-hosted Cybernetic Agent Governance Engine, and HM Treasury’s Financial Services AI Adoption Plan. To test whether their apparent convergence merely reflects a common 2026 attention cycle, the paper adds three retrospective cases: Australia’s Robodebt scheme, Michigan’s MiDAS and claimant-services audits, and the Netherlands’ childcare-benefits scandal.

The paper proposes a six-stage analytical lifecycle: capability, authorization, present admissibility, execution, consequence, and reliance. It asks whether independent reviewers can apply that lifecycle without adopting the author’s terminology or commercial architecture.

The evidence supports further testing of a coherent boundary problem but remains insufficient to establish durable convergence or a mature scientific field. Institutional Computation is therefore presented only as a candidate research programme. A prospective coding protocol, public-review register, explicit disconfirming conditions, and a 31 January 2027 decision point make external use and disagreement—not another author-generated revision—the next required evidence.