To get caught up on this topic, previous posts have covered: an introduction to Agents for National Security, a view on GenAI.mil’s predecessor NIPRGPT, what swarms of agents mean for the DoW.
For the first time, thousands of service members can talk to frontier‑scale models first Gemini, then Claude, Grok, and possbily ChatGPT on government networks, at the IL5 classification levels. Years of policy work, security engineering, and infrastructure building just turned into a login screen and a prompt box.
We’ve already seen a preview of what happens next. When the Air Force rolled out NIPRGPT on NIPRNet, tens of thousands of airmen, guardians, and civilians piled in almost immediately. What started as an “experimental bridge” to generative AI quickly turned into an everyday tool for search, drafting, coding, and analysis.
GenAI.mil takes that same pattern and extends it across the force. We should absolutely celebrate that. It will raise the AI IQ of the institution, normalize responsible experimentation, and surface creative use cases from the bottom up.
But as we all log in and start prompting, we need to be honest with ourselves: chat is the on‑ramp, not the destination. Access to a model via a chat window should be a component of a broader military system to make the DOW faster, smarter, and more efficient.
The hard work ahead is turning this new access into Digital NCOs, Digital Staff Officers, and—eventually—an orchestration layer that can command thousands of those agents across clouds and on‑prem, all wrapped in a resilience layer that keeps them alive in a fight. This is where AI becomes Applied AI, embedded in real workflows that deliver decision superiority and operational impact.
From Chat Windows to Digital NCOs
Right now, GenAI.mil gives everyone in the DoW a “license to browse.” You can ask questions, draft products, refactor code, and get unstuck faster than before. That’s a boost for individual productivity and for learning what these models are actually good at in a military context.
But nobody plans an operation or runs a squadron in a single text box.
In the human world, an NCO is the person who takes a commander’s intent, turns it into real work, and makes sure that work actually gets done. They read the order, pull in the right people, break the problem into tasks, track execution, and report back when something is off. They are the connective tissue between strategy and outcomes.
A bare model sitting behind a chat window can assist with pieces of that, but it can’t own it.
A Digital NCO can start to. Imagine a maintenance Digital NCO that takes “I need this fleet at 90% FMC in 30 days,” pulls maintenance history, parts availability, and schedules, and then proposes a plan: which tail should be worked when, by whom, and with what impact on readiness. Or a digital staff officer that ingests a new OPORD, decomposes it into explicit and implied tasks, opens the task trackers, drafts coordination notes, and keeps an always‑current view of who is red, amber, or green.
In both cases, the human commander and staff are still in charge. But instead of spending their time pushing data between systems, they spend it making decisions. To get there, we need more than just powerful models. We need an architecture built around three core layers: Intelligence, Orchestration, and Resilience.
The Intelligence Layer: From Intent to Structured Work
The Intelligence Layer is where a Digital NCO really “lives.”
This layer understands the mission, the data, and the authorities. It doesn’t just see a blob of text; it sees an OPORD, a FRAGORD, a policy memo, or a maintenance directive and can translate that into structured tasks, constraints, and priorities. It knows where the truth is supposed to live—logistics systems, maintenance databases, readiness tools, personnel records—and can work over those sources instead of relying solely on whatever the model was trained on.
It also knows the difference between what can be done on an unclassified network and what must be handled in a classified enclave. It can choose the right model and environment based on classification, policy, and risk, not just convenience.
Most importantly, the Intelligence Layer doesn’t leave its work trapped in a chat log. When a Digital NCO builds a maintenance plan or updates a readiness roll‑up, those changes flow back into the systems commanders already trust. The unit doesn’t have to move into a chat interface to benefit; the AI comes to where the work already happens.
The Orchestration Layer: Command and Control for AI Agents
If the Intelligence Layer is how a single Digital NCO understands the mission, the Orchestration Layer is how thousands of those Digital NCOs are tasked, coordinated, and supervised.
We’ve just watched the offensive side of this pattern show up in public. Anthropic recently described a state‑sponsored espionage campaign where an attacker used Claude’s “agentic” capabilities to automate a large fraction of the intrusion workflow—tool use, reconnaissance, exploitation—at unprecedented scale.
Analysts writing about the so‑called Factory.ai incident have painted a similar picture in a different domain: AI agents chained together to pursue a sprawling, multi‑platform fraud campaign by stitching free tiers and tools into a single automated scheme.
In both cases, what stands out is not just the intelligence of any single model, but the orchestration: an AI‑driven operations center quietly delegating thousands of small tasks to specialized agents with specific tools and permissions.
That same pattern is exactly what the Department will need—but pointed in the opposite direction, under a clear chain of command.
An orchestration layer for the Department of War should be able to look across clouds, on‑prem data centers, and edge devices and decide which Digital NCO should do what. Some agents may live behind GenAI.mil in a hyperscale cloud; others will sit next to classified data in a JWICS or SIPR enclave; others will run on ruggedized hardware at the tactical edge. The orchestrator’s job is to know where each one is, what it’s allowed to do, and how to sequence their actions into something useful.
Crucially, not every Digital NCO should have the same level of autonomy. Some may be read‑only, allowed to summarize and suggest but never to change data. Others might be trusted to update certain fields, kick off a report, or call specific tools without a human in the loop. A smaller number might be permitted to execute end‑to‑end workflows, as long as their actions are highly observable and easy to override.
The orchestration layer is where all of that is decided and enforced. It’s how the Department of War, in the broad historical sense, will manage a digital corps of NCOs at scale without losing control of what the software is doing on its behalf.
The Resilience Layer: Compute Where the Fight Is
There is one more reality we can’t wish away: the Department of War does not fight in a clean, high‑bandwidth cloud.
GenAI.mil quite rightly leans on hyperscale infrastructure. In garrison and enterprise settings, that is exactly what we want. Frontier models are phenomenal at chewing through massive logistics datasets, generating options, and helping planners explore courses of action they wouldn’t have time to consider otherwise.
But adversaries will go after those links. In a contested environment, relying entirely on distant data centers becomes a single point of failure.
That’s where the Resilience Layer comes in.
This layer treats compute the way commanders treat maneuver forces. In garrison, the Digital NCOs are backed by the biggest models in the cloud. In a SCIF, they may run on models tuned for a classified enclave. At the tactical edge, when a satellite link is flaky or gone, the same workflows fall back to smaller, quantized models sitting on local hardware.
The goal is not to pretend a laptop at a forward operating base can match a hyperscale cluster. The goal is to design workflows that degrade gracefully instead of failing hard. A Digital NCO should keep working—maybe with fewer tools, maybe with more human checks—but it should not disappear the moment the connection gets noisy.
When the Intelligence, Orchestration, and Resilience layers are designed together, commanders see one coherent system: Digital NCOs that understand the mission, are coordinated like units, and keep operating even when the network is anything but friendly.
The Way Ahead
GenAI.mil is an important start. It gives the Department a common, secure platform to experiment with frontier models, learn quickly, and build intuition about what this technology can and can’t do.
But it is just that: a start.
The next phase is about turning the best of those chat‑based experiments into Digital NCOs and Digital Staff Officers embedded in real systems and workflows. Above them, we will need orchestration layers that can safely command thousands of these agents across clouds and on‑prem environments, with clear lines of authority and strong guardrails. And beneath them, we will need resilience layers that make sure all of this survives the conditions we actually train for: contested, degraded, and denied.
The future isn’t one model winning everything. It’s an ecosystem—Gemini, Claude, Grok, service‑specific and open models—operating together under an architecture that understands the mission, respects classification and policy, and works wherever the fight happens.
GenAI.mil has given the force its license to browse. Now we have to build the license to act.