TEOW-AGL: A Pre-Execution Governance Layer for Agentic AI Systems

Agentic artificial intelligence systems can now execute actions in the external world before human operators or governance layers have an opportunity to intervene. Existing safety approaches — prompt guardrails, output filters, permission lists, and post-hoc monitoring — typically operate around the model rather than directly between the model and the tools it commands. TEOW-AGL (Teow Agent Governance Layer) is a model-agnostic pre-execution governance architecture that translates each proposed agent action into a structured AgentAction, evaluates it through a deterministic governance pipeline, and routes it into one of three flows: Blue for autonomous execution, Green for human-gated execution, or Red for emergency blocking. The reference implementation enforces a hard architectural separation between governance authority, human approval, and execution under a single-sequencer orchestrator, with every decision recorded as a JSONL audit trace. The artifact is deliberately scoped as a reference release rather than a production safety platform or compliance solution; its contribution is a reproducible architectural pattern for placing authorization before action. The paper situates this pattern relative to recent work on tool-using language model agents, alignment-side safety methods, and emerging tool-execution standards, and discusses both its capabilities and its explicit boundaries. Companion software: 10.5281/zenodo.19964463 (GitHub: github.com/ai-governance-os/teow-agl).