Imagine an AI\u00a0logistics\u00a0agent for the Department of\u00a0Defense\u00a0that could, in seconds, reroute a fuel convoy around a contested chokepoint, file the necessary clearances, and notify the affected personnel. Imagine a Veterans Affairs AI assistant that walks a Gold Star family through benefits applications without leaving them on hold for\u00a090\u00a0minutes. Imagine a Treasury AI system that catches a billion-dollar daycare scheme before the money leaves the account.<\/p>\n
The promise of agentic AI is massive productivity: faster services for citizens, smarter logistics for warfighters, lower costs for taxpayers, and decision advantage over adversaries who are racing to deploy these same capabilities.<\/p>\n
The problem is we cannot currently deploy agentic AI at\u00a0scale, because today\u2019s systems are not yet reliable enough.<\/p>\n
I spent years at the Pentagon\u2019s Joint Artificial Intelligence Center, helping to\u00a0build the foundations of\u00a0the\u00a0Defense Department\u2019s\u00a0AI enterprise. I watched promising prototypes collide with operational reality. Models can perform brilliantly in demonstrations and then fail in unexpected ways in the real world. In operational contexts, model hallucinations or an adversary\u2019s prompt injection attack could be catastrophic. Warfighters need to be able to trust that an AI is not\u00a0optimizing for\u00a0the wrong objective. We need to ensure that the AI system behaves the same way in testing and production.<\/p>\n
Unlike normal software, AI systems are not programmed by\u00a0a human. They write their own code and decide how they interpret operator instructions. When an AI system goes wrong,\u00a0there\u2019s\u00a0no stack trace you can examine. Guiding these AI systems reliably is the central engineering reality of frontier AI in 2026. It is unlike other software challenges the government has faced. And the higher the stakes of the deployment, and the more capable the AI system, the less acceptable these mistakes are.<\/p>\n
A chatbot that gets a fact wrong is an inconvenience. A superintelligent agentic system with access to classified military data, financial controls or critical infrastructure is a national security incident waiting to happen.<\/p>\n
I have spent the last two years working with lawmakers on what I believe is the central AI policy challenge of this Congress: building the science, the testing frameworks and standards, and the institutions that will unlock trustworthy agentic\u00a0artificial intelligence\u00a0that our nation can count on when the stakes are highest.<\/p>\n
A bipartisan coalition of national security and AI experts has urged Congress to fund the National AI Reliability and Control Initiative (NAIRCI) at\u00a0$2 billion in the\u00a0fiscal\u00a02027 National Defense Authorization Act. NAIRCI provides targeted research funding for the hardest unsolved problems in AI: how to make systems behave predictably, how to verify they are doing what we asked, how to keep them aligned with human intent as their capabilities expand, and how to keep meaningful human stewardship over these systems. This is the science that will accelerate American AI companies, defense\u00a0primes\u00a0and businesses in their missions to deliver products that our government can\u00a0actually field.<\/p>\n
Some will argue that any guardrails will slow America down in our competition with China. The opposite is true. Technical AI safety is loadbearing for American AI dominance.<\/p>\n
The People\u2019s Republic is racing to deploy agentic AI inside its own state apparatus, and Chinese researchers are\u00a0encountering\u00a0the same reliability problems we are.\u00a0The country that solves them first will not merely have safer AI; it\u00a0will have AI that actually works at scale in the systems that matter most.\u00a0Reliability is not a tax on innovation. It is the precondition for the kind of innovation that ends in real deployment rather than fancy marketing, failed\u00a0pilots\u00a0and stalled procurements.<\/p>\n
Finally, I would note that we are in two races at once. The first is the commercial and military AI competition with the People\u2019s Republic of China, which we must win. Market share is the goal. The second is the longer race, against time itself, to ensure that increasingly powerful systems remain under human stewardship as their capabilities grow. Both races are won the same way: by investing in the science of reliability, by building the evaluation infrastructure, and by creating the policy frameworks that let American companies and American agencies accelerate the deployment of trustworthy AI capabilities.<\/p>\n
Agentic AI will transform the government. The question is whether that transformation will be one we shape deliberately, with the safeguards a free society\u00a0requires, or one that happens to us in the wake of a cascading failure no one\u00a0anticipated.<\/p>\n
Congress has a clear mandate. The American public, in poll after poll, has made clear that they want both the benefits of AI and assurance that the technology is being deployed responsibly. The U.S. technical community is the best in the world and stands by to\u00a0assist. What\u00a0remains\u00a0is the will to act.<\/p>\n
Mark Beall is\u00a0president of the AI Policy Network. He served as the inaugural\u00a0director of AI Strategy and Policy at the Pentagon\u2019s Joint Artificial Intelligence Center and has testified before Congress on AI and national security.<\/p>\n
Copyright 2026 Nexstar Media, Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.<\/p>\n