The Pentagon Built 100,000 AI Agents. Most Workers Havent Learned How to Run Them.

The Pentagon has built more than 103,000 AI agents in under five weeks — and as of this week, only 26,000 of the three million workers on the GenAI.mil platform have completed any formal AI training. An AI agent, in this context, is a software program that runs a sequence of tasks autonomously, like drafting a procurement document, checking a regulation, or summarizing a policy memo. The no-code tool workers used to build them requires no engineering background and no certification — which is exactly how 103,000 of them got built before anyone finished the optional training courses.

The training gap is the part of the AI adoption story that nobody puts in the press release. The courses are oversubscribed. The platform is not. Of the 1.2 million unique users who had tried Gemini chat as of early March, only 26,000 had finished formal AI courses, according to AI CERTs News. Those figures have aged since March; what is new this week is the agent count, reported by Breaking Defense.

What the announcements did not say is that the agents exist in a constrained operating environment that Google and the Defense Department deliberately built. Memory features are disabled until compliance reviews finish. Every agent execution requires a human review before anything runs in a production context. The architecture is not optional caution layered onto a fast rollout. It is the guardrail that makes the fast rollout politically and legally survivable. The question is whether the guardrails hold when the humans doing the reviewing have not been trained to review what they are looking at.

Emil Michael, the Defense Department's chief technology officer, called the Agent Designer launch — a tool that debuted on GenAI.mil in early March 2026 — a confidence vote for Google Gemini, according to DefenseScoop. The platform reached one million unique users within a month of its December 2025 launch. The Pentagon cut ties with Anthropic in early 2026 after a dispute over the company's terms of service for autonomous weapons and mass surveillance use cases; Anthropic filed suit challenging the supply chain risk designation. The legal fight is unresolved.

The vibe-coding label — building with AI tools without formal engineering training — is the most accurate description of what the Pentagon workforce has been doing. Non-engineers are now the development force. The bottleneck on AI deployment inside the Defense Department is no longer whether the tools exist or whether they work. It is whether the institution can develop the judgment to oversee what it has already built.

That is not a problem unique to the Pentagon. Every large organization that has moved from AI-as-chatbot to AI-as-agent is running into the same wall, just with smaller numbers and less public accountability. The Defense Department is the most visible test case because it is the largest, because it deals in classified information, and because its failures have consequences that are not hypothetical.

What happens when the 103,000 agents include compliance-checking workflows that hallucinate a regulation, budget agents that misread a continuing resolution, or logistics agents that chain a misconfigured connector to an actual procurement system? The human review requirement is supposed to catch those errors. Whether the humans doing the reviewing know what they are looking at is a different question.

The classified network expansion that Pentagon leadership has discussed as a future milestone is on hold pending the unclassified results. The results, in the form of 103,000 agents and four million uploaded documents, are already there. The oversight infrastructure is still being built. The order of operations got reversed. The experiment ran. The conclusions will have to catch up.