One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI
Source
Gaston Besanson
Author
Gaston Besanson
Date
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Why it matters
If your AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → runs multiple pre-action gates, a fix applied by one can silently invalidate another's approval — and remediation order changes the outcome, so it belongs in your control-plane spec rather than in whatever order the code happens to run.
articleSARC-DQ: Runtime Data-Quality Gating for Agentic AI: Silent Evidence Defects, the Incompetence Shield, and Downstream-Only RemediationGaston Besanson
articleTesting and Evaluation of Agentic AI Systems In Military Command and ControlUlysse Richard, Heather Frase, Sarah Cao, Di Cooke, Sebastian Kwon, Adrianna Tan
articleSpecification-first convergence with an AI coding agent: a case study of dismantling a core architectural invariant across 189 files in a 717k-line codebase with no test oracle and no human code reviewJoel Abenhaim