Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide.
On Sept. 7, 2025, Ukraine intercepted most of the 800 drones and missiles Russia launched in a single, rolling barrage. But enough got through. The lesson wasn’t about any single weapon’s sophistication. It was about mass, and what mass does to tempo when it arrives in swarms. Wars now bend by overwhelming the decision loop.
The same dynamic is coming for government operations. The “airframes” will be AI agents — specialized software workers that observe, analyze, decide, and act within defined parameters. A thousand agents compress the OODA loop — observe, orient, decide, act — until adversaries are perpetually reacting.
That’s the game: tempo.
Yet the U.S. government is fixated on the wrong layer. Agencies just signed dollar-menu AI deals — $1-a-year access here, 47 cents there — followed by inevitable protests. Buried in the coverage: Some bargain tiers exclude APIs. You can type in a window but can’t wire it into workflows. The General Services Administration launched USAi, a government-run sandbox for experimenting with commercial models. The Air Force fielded NIPRGPT on unclassified networks. Then the Army blocked access over data governance concerns.
As the CEO of a software company that offers AI solutions for the Defense Department, I have a commercial stake in this debate. However, this experience — along with my doctoral research and my time at RAND — has given me a unique perspective on these issues.
So, in my view, these government AI initiatives represent the first five minutes of the story. The mission question is plainer: How can the U.S. government reclaim time from administrative burden and accelerate decision cycles?
Start practical. Aim to reclaim 10-to-20 percent of the administrative day by deploying agents for repetitive tasks: forms that file themselves, tickets that propose mitigations with evidence attached, logistics requests that draft, route, and audit with full provenance. By 2030, mainstream research suggests up to 30 percent of work hours could be automated. Take the conservative estimate. Apply it to workflows that consume weeks.
Those reclaimed hours accelerate operational tempo. Intelligence agents monitor approved feeds, attach provenance and confidence scores, and propose courses of action with risk assessments. Humans decide; execution agents implement and log. Command and control agents maintain the common operating picture, escalating only threshold-crossing changes while pre-positioning options. Logistics agents crawl inventories, draft requisitions, and book transport. Cyber agents baseline configurations, detect anomalies, propose containment with blast-radius estimates, and execute post-approval.
Making this work requires an agent control plane — traffic control for software workers. Every agent needs identity. Every tool call needs least-privilege policy. The architecture must include cross-enclave scheduling, observability like flight recorders, tool brokers with allow-lists, data gateways with row-level entitlements and lineage tags. And yes, a kill switch.
With this foundation, the vendor conversation becomes manageable: route each task to the appropriate model — commercial here, open-source there — without rewiring infrastructure. Contracts that exclude APIs should be non-starters. USAi transforms from a bot gallery into connective tissue for agents. NIPRGPT’s lessons fold into a control plane that works across services.
The open-source-versus-commercial debate misses the point. The antidote to vendor lock-in is owning the orchestration layer and audit trail. Control the interfaces to control the leverage. Vendors will evolve. Missions persist.
Return to Ukraine’s lesson: Saturation is doctrine. The battlefield became a traffic jam because one side decided volume would steal time from its adversary. America’s answer should be fielding the swarm to pull U.S. decision loops ahead of competitors.
The difference between assistants and agents is the difference between theater and effects. Assistants help humans type faster. Agents complete missions. They execute bounded tasks, report results, and move to the next objective. They transform tempo.
Consider a concrete example: processing security clearance updates. Today, analysts spend hours cross-referencing databases, filling forms, routing approvals. An agent swarm could monitor personnel records for triggering events, automatically gather required documentation, pre-fill submissions with cited sources, route to appropriate reviewers based on real-time availability, and track completion metrics. The analyst shifts from paperwork to exception handling and quality control. Multiply this pattern across every administrative function.
The technology exists. Open-source models handle classified data on air-gapped networks. Commercial APIs provide sophisticated reasoning. The control plane architecture is proven in private sector applications. What’s missing is the will to move beyond pilots to production.
While America debates contract terms and sandbox architectures, adversaries compress their loops. The choice is clear: continue admiring chat interfaces or field the agent corps as an operational imperative. In the digital battlespace, as in Ukraine’s skies, whoever controls tempo controls outcomes.
The swarm is coming. The only question is whether America is launching it or defending against it.
Ben Van Roo is the co-founder and CEO of Legion Intelligence, an agentic AI platform built for the Department of Defense. Ben has spent his career building tech companies serving the public and private sectors, including five years as researcher at RAND and a startup executive for 13 years. Ben has a PhD in operations research from the University of Wisconsin-Madison.
**Please note, as a matter of house style War on the Rocks will not use a different name for the U.S. Department of Defense until and unless the name is changed by statute by the U.S. Congress.
Image: Midjourney
