Analyst: Pentagon Must Command AI Forces

A Pentagon poster urging personnel to "use AI" contains a fundamental flaw, according to a military analyst writing for War on the Rocks. The Department of Defense wins by commanding forces, not by using tools, and must apply this principle to artificial intelligence by building military agentic forces under human commanders. Senior leaders should stop viewing AI agents merely as tools that augment humans, the source states. They should also abandon the assumption that closed-weight frontier models are the only technological edge. Instead, the Department of Defense should organize AI agents under commanders, qualify them on real missions, field them on models that units control, and let humans own the operational risk.

The Case for Agentic Command

The author draws on personal experience commanding a hobbyist "task force" of AI agents from a Maryland basement since 2023. This force operates on the mission command philosophy: centralized planning, decentralized execution, and clear commander's intent. The argument is that the military must transition from using AI to commanding it, establishing proper subordination where commanders own all authority and risk. There is no mission command without subordinates and intent. A 2012 paper by the Chairman of the Joint Chiefs of Staff states commanders must understand subordinates to translate clear intent. Ordering a human into a position creates answerability. This relationship does not automatically exist with AI. Therefore, the Department of Defense must ensure agentic command relationships by giving commanders frameworks to understand agents, using that understanding to translate intent, and developing trust by observing operational performance when agents are cut off from command and control.

Qualifying AI Subordinates

An agentic warfighter is defined as a model plus scaffolding and intent. The model provides reasoning and capability. Scaffolding includes tools, memory, and planning loops. Under a human commander with a mission, this becomes an agentic warfighting unit. The source contends this form of warfare is already here and moving quickly. Qualifying these AI subordinates requires looking beyond simple benchmark scores. In 2023, models could complete certain expert tasks only about half the time. Today, their performance horizon has extended. The nonprofit Model Evaluation and Threat Research provides useful raw task results for qualification. These results are uneven. For instance, Anthropic's Claude Mythos Preview solved a complex 30-hour robotics task perfectly but failed a ten-second task that an older model handled. It excelled at writing exploits but failed a cryptanalysis challenge. The source suggests equating these jagged results to task-based certification, feeding into mission-essential task assessments for readiness. For offensive cyber operations, frameworks like Cybench are cited as useful starting points for commanders to understand their AI subordinates' capabilities and limitations. Commanders can use our stats and injuries pages to track performance and operational readiness of both human and AI units.

Implementing Command and Control

The Department of Defense is moving toward greater AI integration but is missing the command element, the analyst argues. A June 2026 national security memorandum explicitly highlighted this gap, ordering a rewrite of Directive 3000.09 on Autonomy in Weapon Systems to ensure such systems "respect the chain of command." Respect requires discretion-the choice to follow or break a rule-which is a quality of subordinates, not tools. Congress is also moving. A draft Senate authorization bill, advanced 18 to nine by the Armed Services Committee, could establish a unified combatant command called the Robotic and Autonomous Systems Command. Whether this specific command is created or not, the source urges the Pentagon to proactively command agents as subordinates. Failure to solve the command problem first risks a scenario where the military cedes protected networks to rogue AI tools. Understanding the broader strategic standings and operational fixtures is crucial for integrating such commands.

The Risks of Remote Cognition

A final test for AI subordinates is the ability to execute commander's intent when cut off from command and control. The Department of Defense's pivot to competition and reliance on closed frontier models skews it toward centralization, which contradicts the decentralized execution half of mission command. Sending data to a vendor's remote cognition platform creates friction and a delay chain vulnerable to adversary disruption. While solutions like running Gemini AI on an air-gapped Google Distributed Cloud exist, data centers that units cannot physically reach still represent remote cognition and are high-value targets. Disruption can also come from policy, as demonstrated by a 2026 order for federal agencies to drop Anthropic models, showing how contracts or vendor conflicts could kill models before an adversary acts. Agents relying on remote cognition answer to both the unit and the vendor, creating a conflict where only one can be the true commander. The source warns that China's military is centralizing its AI, testing whether it can replace commanders its leadership no longer trusts. Adopting remote cognition forfeits the decentralized advantage of mission command and plays into an adversary's desired end state. Commanders who trust their agents to operate within intent when cut off can accept more risk and generate more paths to mission success in a contested environment. This requires a deep understanding of the squad, whether human or AI, to build effective, resilient units.