On January 12, speaking at SpaceX’s Starbase site in Texas, Defense Secretary Pete Hegseth released three memos reshaping how the department develops and adopts technology. One of them, “Artificial Intelligence Strategy for the Department of War,” directs the department to become an “‘AI-first’ warfighting force across all components, from front to back,” according to DefenseScoop.
“We’re done running a peacetime science fair while our adversaries are running a wartime arms race,” Hegseth said, as quoted by Breaking Defense.
/ 01What the strategy says
The six-page AI memo sets out seven “Pace-Setting Projects” to be led by the Chief Digital and AI Office. As described by Breaking Defense, they are:
- GenAI.mil, the generative AI platform launched in December, whose models the memo says will eventually be available “at all classification levels.”
- Swarm Forge, pairing elite units with technology innovators to develop “novel ways of fighting with and against AI-enabled capabilities.”
- Agent Network, semi-autonomous AI agents for battle management and decision support, from campaign planning to kill chain execution.
- Ender’s Foundry, AI-enabled simulation.
- Open Arsenal, linking intelligence collection to capability development.
- Project Grant, described only as aiming to “enable the transformation of deterrence.”
- Enterprise Agents, AI agents for back-office enterprise systems.
DefenseScoop reported further directives: the CDAO must set benchmarks for model objectivity as a primary procurement criterion within 90 days; “any lawful use” language is to be added to AI contracts within 180 days; and appropriate data from weapons and IT systems must be made available for “AI exploitation.” A separate memo realigns the CDAO, DARPA, DIU and other organisations under the department’s chief technology officer, Emil Michael. A third splits the Advana data environment, separating financial data from a new War Data Platform.
On data access, Breaking Defense reported that the department strengthened earlier data decrees: any denial of a CDAO data request must now be justified to the under secretary for research and engineering within seven days.
/ 02Why it matters
The strategy is notable less for its ambition, which previous strategies shared, than for its specificity: named projects, short deadlines, and a mechanism to escalate data refusals. Former officials quoted by Breaking Defense were cautious. Retired Lt. Gen. Jack Shanahan, the founding director of the Joint AI Center, said: “I remain cynical about execution, given how the building operates.”
Execution will hinge on data. Agents for battle management and decision support are only as good as the picture they reason over, and that picture is assembled from many systems that were never designed to share a common vocabulary. The seven-day rule on data requests addresses access. It does not address meaning.
/ 03The knowledge-graph angle
Getting data out of a system is the first step. Making it usable by an agent means reconciling it: recognising that a unit, a location or a piece of equipment described in one system is the same entity described differently in another, and recording how confident that match is and where it came from. Without that, an agent network will produce decisions that depend on silent joins nobody can inspect.
This is where the defense community’s existing work on formal ontologies, including the Basic Formal Ontology and the Common Core Ontologies built on it, becomes practical rather than academic. A shared ontology gives data from different programmes a common structure; a knowledge graph built on it keeps each fact tied to its source; and deterministic reasoning over that graph lets a commander or reviewer see why an agent recommended what it did. For decision support “from campaign planning to kill chain execution,” that audit trail is not a nicety.
Speed and traceability are often framed as a trade-off. In practice, the reverse is usually true at scale. Every time an agent’s recommendation cannot be explained, someone has to re-derive it by hand before acting, and that is where time is lost.
Speed in AI adoption comes from shared meaning in the data, not just shared access to it.
/ 04What to watch
- The 90-day benchmarks for model objectivity, and how they are measured.
- How the War Data Platform is structured, and whether it adopts a common semantic model across sources.
- Agent Network’s requirements for explainability and human review as it moves toward operational use.
The seven-day rule forces data out of programme silos. It does nothing to make that data agree, and Agent Network will be supporting campaign planning on top of whatever joins get made in a hurry.
A course of action is a chain of dependencies: tasks need capabilities, capabilities are provided by units, and units have a reported location and readiness from a particular system at a particular time. Built on BFO and the Common Core Ontologies, a graph keeps each link with its source and timestamp, so when a logistics feed and an intelligence report disagree about where a unit is, the conflict is recorded rather than averaged away. A planner can then ask which assumptions in a recommended course of action rest on stale data or on a single source, and get a list with citations. The regional-conflict build on our site fused six sources into 720 records, each traceable to its document; that property is what an agent network needs, inside the perimeter.
How we’d model it
Course of action—requires→CapabilityCapability—providedBy→UnitUnit—reportedAt→LocationLocation report—sourcedFrom→Data system
- 'Accelerate like hell': Hegseth moves to reshape DOD's AI and tech hubs — DefenseScoop defensescoop.com ↗
- Pentagon rolls out major reforms of R&D, AI — Breaking Defense breakingdefense.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.