Military agent models and governance

A new military artificial intelligence product called WarClaw was released on April 1 by EdgeRunner AI, and its selling point is almost the opposite of consumer chatbots: it runs locally on laptops and servers, works without the internet, and is built for denied, disconnected, intermittent, and low-bandwidth conditions instead of cloud convenience. (edgerunnerai.com) That matters because front-line units often lose clean connectivity first, not last, and EdgeRunner says WarClaw is meant to keep working when troops are cut off from the cloud. The company describes it as a “hardened agentic orchestration layer” that can search databases, interpret intelligence reports, draft briefings, and automate routine staff work on-device. (edgerunnerai.com) The bigger shift is not one product but the kind of product. Defense One reported on April 1 that WarClaw was trained by former operators and military subject-matter experts on actual military tasks, which puts it in a different category from general-purpose frontier models built for the public internet. (defenseone.com) An agent is software that does more than answer a question once. It can take a goal like “build the morning intelligence brief,” pull files, sort evidence, write drafts, and hand a package to a human the way a junior staff officer would, except at machine speed. (defenseone.com) The Pentagon is already moving in that direction. In its January 12, 2026 Artificial Intelligence Acceleration Strategy, the Department of War said it is building an “Agent Network” for battle management and decision support “from campaign planning to kill chain execution,” alongside a separate “Enterprise Agents” effort for office workflows. (war.gov) That is where the governance fight starts, because the main U.S. rulebook was written for weapons, not for recommendation engines that shape human choices before a shot is fired. Department of Defense Directive 3000.09, updated on January 25, 2023, covers autonomous and semi-autonomous weapon systems and is designed to reduce failures that could lead to unintended engagements. (esd.whs.mil) The directive draws lines based on the human role. The Congressional Research Service says it distinguishes between fully autonomous systems that can select and engage targets without further human intervention, human-supervised systems that can be monitored and halted, and semi-autonomous systems that engage targets selected by a human operator. (congress.gov) But many of the newest military artificial intelligence tools are not pulling the trigger themselves. The Institute for AI Policy and Strategy wrote on April 9 that artificial intelligence decision-support systems can analyze intelligence, generate target recommendations, and rank courses of action while still falling outside Directive 3000.09. (iaps.ai) That gap is easy to miss if you picture a robot weapon and not a recommendation screen. A commander may still be the formal decision-maker, but if software has already filtered the data, highlighted one target, and compressed the timeline, the machine has shaped the decision before the human arrives. (iaps.ai) The policy memo from the Institute for AI Policy and Strategy says these systems can fail in very old ways and very new ways at once: bad data, stale data, adversarial manipulation, and model misalignment can all flow into use-of-force decisions. The memo also says some military decision-support systems rely on commercial frontier models that may not have been tested enough for security and reliability in combat settings. (iaps.ai) That is why purpose-built systems like WarClaw are getting attention. EdgeRunner says the tool is designed for operator oversight and local deployment, while recent Pentagon guidance on testing says artificial intelligence systems need evaluation for calibrated trust, emergent behavior, human-machine teaming, and responsible artificial intelligence compliance before they are relied on in the field. (edgerunnerai.com, aaf.dau.edu) The recommendations now on the table are not abstract. The Institute for AI Policy and Strategy calls for four concrete fixes: meaningful human accountability, research and development for reliability and security, acquisition standards for reliability and monitorability, and training for operators, commanders, and technical staff. (iaps.ai) The Pentagon’s own responsible artificial intelligence pathway points the same way. Its guidance says trust has to be earned during design, testing, procurement, deployment, and use, not added after launch like a warning label on a box. (media.defense.gov)