Japanese overview of the proposed pre-execution controls.
AI AGENT CONTROL / PUBLIC EXPLANATION
How to stop an AI agent before an external action
Input filtering alone is not a sufficient stop point. The Japanese source proposes checking request structure, evidence, permission and approval just before an agent sends, orders or publishes.
Why entry filters are insufficient
Filters judging mainly meaning can be evaded by changes in wording or context. A more useful stop point is directly before the proposed action affects an external system.
- Design aim: retain a clear opportunity to hold an external action.
Examine structure and evidence
Topological Gatekeeper (TG) examines the structure, path and deviations of a request, rather than relying only on what it says. Evidence Adequacy Gate (EAG) checks whether grounds for execution or publication are sufficient. Rephrasing cannot by itself repair missing approval or evidence.
- Design aim: reduce dependence on surface phrasing for execution decisions.
HOLD is a normal operating state
The source explains PASS, HOLD and ESCALATE as conceptual outcomes. HOLD means information, authority or a safety condition remains unresolved, so the proposed action stops for now rather than being treated as a mere system error.
- This differs from the ITS API's HOLD / ALLOW / DENY terminology.
Related sources and materials
Japanese structural-gate specification.
PoC-stage action-gate API and its limits.
Demonstration, not production enforcement.
Conditions for authorized external actions.
doc_id: C3-WEB-ANSWERS-STOP-AI-AGENTS-EN-0.1 / version: 0.1.0 / status: REVIEW_REQUIRED / translated: 2026-10-13 / source_doc_id: C3-WEB-ANSWERS-STOP-AI-AGENTS-0.1 / source_last_updated: 2026-06-05