U.S. Marine Corps / Sgt. Raymond Tong

A new agentic-AI tool will continuously scan intelligence feeds and operational networks to provide U.S. military commanders with targeting options “within seconds,” the Pentagon announced Thursday.

Dubbed Agent Network, the new tool will employ “agents”—artificial-intelligence entities that perform tasks on behalf of a user, such as running a scheduled search or executing an email campaign—to “continuously scan defense intelligence and operational systems, translating findings into clearly presented options,” said a press release, which added: “Agent Network does not autonomously select or strike targets; it ensures commanders remain in charge of every decision.”

It is one of seven “pace-setting” projects originally unveiled in January along with a new Pentagon AI strategy. Key contractors in the Agent Network effort include Lumbra and Palantir, which already handles much targeting analysis through its Maven Smart Systems contract.

But expectations for what agents can currently do may be running ahead of reality. “Tasks that AI agents are instructed to perform can clearly have computational complexity beyond” what current large language model architectures can handle, Vishal Sikka—a former CEO of SAP—wrote last July.

Citing the seminal Time-Hierarchy Theorem, Sikka noted that transformer models approach difficult tasks and simple ones using the same mechanical formula. These models can only perform so many operations per “token,” which is the way large language models understand word concepts. Even dealing with seemingly simple concepts can require a large number of tokens. Because of this limitation, there is no way to get a transformer-based model to not hallucinate when the task that you give it is more complex than it has tokens to bring to that task.

“Despite their obvious power and applicability in various domains, extreme care must be used before applying LLMs to problems or use cases that require accuracy, or solving problems of non-trivial complexity,” Sikka concluded.

But Illia Pashkov, founder of SINT Labs and editor of The Agent Times, cautioned against underestimating agents’ potential. 

“Agentic AI quietly stopped being a demo this year,” Pashkov said. “It's drafting code, clearing support queues, grinding through back-office work in finance and healthcare, and now it's reading intelligence. The speed is not hype. I've watched these systems compress weeks of analyst work into an afternoon.”

But their capabilities also bring risks—more than people accustomed to working with common AI chatbots might realize. Private-sector companies that have rushed to put AI agents to work are already seeing problems, Pashkov said, pointing to a company whose agent wiped a live production database. Unless carefully implemented, agents can’t tell when they’re going wrong.

“The danger was never a dumb agent; it's a confident one running without a leash, a logbook, or a human who owns the call,” he said. 

A host of Defense Department offices and teams are beginning to deploy agent systems, said one DOD intelligence security official who is not directly affiliated with the Agent Network program. The official described an atmosphere of enthusiasm. 

“There are so many opportunities to leverage the DOD Enterprise capabilities and allow people to build their own agents,” they said.

But the official allowed that keeping track of how every agent is performing is a major challenge. Governing all of them will be nearly impossible.