Agentic Warfare

The shift from manual, human-driven military decision cycles to AI agent-driven operations. It is the defense application of the same infrastructure and methodology reshaping enterprise workflows: just as corporate processes are being converted into agent-driven pipelines, military planning processes are being converted into AI agent workflows, with decision-making cycles compressing from days to minutes.

The current state of military decision-making

Alexandr Wang's framing: "If you mapped out what warfare looks like today, Russia and Ukraine or other conflict areas, the decision-making processes are remarkably manual and human-driven. Very critical battle-time decisions are made with very limited information in very manual workflows."¹

This is the same pattern as corporate enterprise workflows: humans clicking around, aggregating information from multiple sources, producing analysis, making decisions, all slow and lossy. The structural gap between potential information quality and actual decision quality is enormous.

The agentic warfare thesis

AI agents applied to military decision cycles enable:

• Perfect information: agents can aggregate and synthesize across all available data in real time
• Immediate decision cycles: no human bottleneck on information processing
• Doctrinal compliance: agents can encode existing military planning doctrine and execute it rather than deviate from it

The result, in Wang's words, is "almost incomprehensibly fast-moving scenarios."¹ His chess analogy is precise: playing chess against a human requires waiting for them to think. Playing against a computer means immediate responses, unrelenting. Agentic warfare transforms military conflict the same way.

Thunder Forge

Scale AI's flagship Department of Defense program, built with the Indo-Pacific Command in Hawaii, responsible for the Indo-Pacific region including China, the Taiwan Strait, and the South China Sea.

Thunder Forge converts the existing human-driven doctrinal military planning process into a series of AI agents that work together to conduct the same planning task. The result: 72-hour decision cycles compress to 10 minutes.¹

This is explicitly the most consequential agent workflow in Scale's portfolio, not because it is technically the hardest, but because the stakes are highest. The same pattern used in hiring, deep research plus synthesis, applies to operational planning.

Fragmentation as the broader trend

Wang's geopolitical observation is that the Cold War model, bigger and bigger bombs, deterrence through escalation, is being replaced by fragmentation. The new defense paradigm favors drones and embodied robots, cyber warfare, and smaller, more nimble, attributable resources: "the exact opposite of that Cold War philosophy. It's the fragmentation and the move toward smaller, more nimble, attributable resources."¹

This makes agentic decision-making more critical, not less. The complexity of coordinating many small autonomous assets exceeds human cognitive capacity, and the agent layer becomes the coordination layer.

Connection to competition with China

China's manufacturing advantage in robotics hardware, roughly two to four thousand dollars per unit against twenty to thirty thousand in the United States, combined with its data collection infrastructure, government-run labeling centers and robotics data factories, creates a potentially insurmountable advantage in the physical embodied AI layer of warfare. Wang: "The degree to which China is incredible at manufacturing is a very big problem. It relates to defense and national security."¹

The agentic warfare thesis compounds this: if decision cycles are compressed and the hardware is cheaply producible, the country that can mass-produce the hardware and run it at machine speed wins.

Open questions

Thunder Forge's actual efficacy is unverified since the program is classified; the claimed compression from 72 hours to 10 minutes is asserted, not independently confirmed. Perfect information is an aspiration rather than a technical reality, since agent errors compound at speed. And the incomprehensibly fast-moving scenario raises arms control questions: if both sides run agentic warfare systems, escalation dynamics could become autonomous and faster than human intervention can interrupt.