Google’s AI Agent Tool Gets a Huge Pentagon Win, Pressuring Microsoft

TL;DR

• Defense Department personnel built over 103,000 semi-autonomous AI agents using Google Gemini’s Agent Designer tool on GenAI.mil in under five weeks.
• The platform recorded 1.1 million total agent sessions, with 180,000 sessions happening weekly as of mid-April 2026.
• GenAI.mil enables DoD workers to build custom AI agents on unclassified networks without traditional coding skills.
• The numbers signal Google’s growing enterprise AI traction in defense, competing directly against Microsoft and AWS for military contracts.

DoD Workers Built 103,000 Agents in Weeks

Pentagon personnel created more than 103,000 semi-autonomous AI agents in under five weeks using Google Gemini’s Agent Designer tool, according to a Defense Department official speaking to Breaking Defense. The agents run on GenAI.mil, the Pentagon’s platform for building custom AI tools on unclassified networks. “We’ve seen remarkable adoption since its launch, with over 103,000 agents built and a total of more than 1.1 million agent sessions recorded” as of mid-April on GenAI.mil, the official said.

The platform now sees roughly 180,000 agent sessions per week. That’s not just experimentation — it’s operational tempo. These agents handle tasks ranging from data analysis to workflow automation, all without requiring DoD personnel to write traditional code.

Google’s Agent Designer lets users describe what they want an AI agent to do in plain language. The tool then generates the agent. It’s the enterprise version of what consumer AI promised but rarely delivered: automation that doesn’t require a computer science degree.

Why Google’s Gemini Win Matters for Defense AI

This isn’t just a big number. It’s a signal that agentic AI — systems that can act semi-autonomously on tasks rather than just answer questions — has crossed into military operations faster than most expected. The Pentagon doesn’t typically move this fast on new technology adoption. Five weeks from launch to six-figure agent creation is breakneck for an institution that still runs some systems on COBOL.

And the competitive stakes are real. Google’s landing this kind of traction inside the Defense Department puts it in direct competition with Microsoft, which has been pushing its Azure OpenAI services aggressively into government contracts, and AWS, which dominates cloud infrastructure for defense. Microsoft’s got the Office 365 install base and the OpenAI partnership. AWS has the security clearances and the infrastructure footprint. Google’s now got the agent-building platform that DoD workers actually use at scale.

I’ll admit, I didn’t expect Google to outflank Microsoft here. Google’s enterprise AI strategy has felt reactive compared to Microsoft’s Copilot blitz across every product line. But Agent Designer on GenAI.mil is a different play — it’s not about embedding AI into existing tools. It’s about letting users build their own tools. That’s a fundamentally more flexible approach, and the adoption numbers suggest it resonates with how defense personnel actually want to use AI.

Think of it like the difference between getting a Swiss Army knife and getting access to a workshop. Microsoft’s handing out knives — useful, pre-configured, limited. Google’s handing out the workshop. The Pentagon apparently wants the workshop.

The 1.1 million sessions number is worth unpacking. That’s not 1.1 million agents — it’s 1.1 million times someone used an agent. If you’ve got 103,000 agents and 1.1 million sessions in five weeks, that’s an average of about 10 sessions per agent. Some are probably one-off experiments. But many are clearly getting repeated use. That’s the difference between a demo and a tool.

GenAI.mil and the Unclassified Network Advantage

GenAI.mil exists because the Pentagon needed a way to let personnel experiment with generative AI without waiting for classified system approvals. It runs on unclassified networks, which means it can’t touch sensitive data — but it also means it doesn’t require the months-long security reviews that classified AI systems demand. Speed matters when the technology is evolving this fast.

The platform’s design reflects a broader shift in how the military thinks about AI adoption. Instead of top-down deployment of a single AI system, GenAI.mil enables bottom-up experimentation. Individual analysts, logistics officers, and administrative personnel can build agents tailored to their specific workflows. That’s how you get to 103,000 agents in five weeks — you’re not building one system for everyone, you’re letting everyone build their own.

This approach has risks. Decentralized AI agent creation means less control over what those agents do and how they’re trained. The Pentagon’s betting that the speed and flexibility gains outweigh the governance headaches. So far, the adoption numbers suggest users agree.

But the unclassified-only constraint is real. These agents can’t access the data that matters most for national security work. They’re useful for administrative tasks, unclassified research, logistics planning, and workflow automation — valuable, but not the core mission. The question is whether the Pentagon can replicate this adoption rate when it eventually rolls out agent-building tools on classified networks. That’s a much harder technical and bureaucratic problem.

What the 180,000 Weekly Sessions Signal

The 180,000 weekly sessions figure tells you this isn’t slowing down. If anything, it’s accelerating. That’s the current run rate as of mid-April, which means usage has likely grown since the 1.1 million total session count was recorded. People are building agents, using them, iterating on them, and building more.

What are they building? The Defense Department hasn’t released specifics, but the use cases likely cluster around data wrangling, report generation, meeting summarization, and workflow automation — the unglamorous but time-consuming tasks that eat up hours of knowledge worker time. If an agent can take a pile of unstructured data and turn it into a formatted briefing slide, that’s a win. If it can monitor a data feed and flag anomalies, that’s a win. These aren’t science fiction use cases. They’re just automation that finally works without requiring a developer.

The broader implication is that the military is learning how to use agentic AI in production, at scale, right now. That’s a strategic advantage. While other governments are still running pilots and writing white papers, the Pentagon is racking up operational experience. When classified agent platforms eventually come online, DoD personnel won’t be starting from zero — they’ll already know how to design, deploy, and manage AI agents. That’s the real win here.

Google’s Defense Traction and What Comes Next

For Google, this is validation that its enterprise AI strategy can compete with Microsoft’s in a high-stakes market. Defense contracts are lucrative, but they’re also reputation-builders. If Google can point to six-figure agent adoption in the Pentagon, that’s a selling point for every other government agency and Fortune 500 CIO evaluating AI platforms.

The question is whether Google can sustain this lead. Microsoft isn’t going to cede the government AI market without a fight, and it’s got deep relationships across federal agencies. AWS has the infrastructure lock-in. Google’s advantage right now is that Agent Designer apparently nails the user experience in a way that resonates with how defense personnel want to work. But user experience advantages are fragile — they can be copied, and they can be overcome by ecosystem lock-in.

Watch whether Google expands GenAI.mil’s capabilities to include more complex agent orchestration — multi-agent systems that can coordinate with each other. Watch whether Microsoft counters with its own agent-building platform tailored for government use. And watch whether the Pentagon starts publishing more detailed metrics on what these agents actually do and how much time they save. Right now, we’ve got adoption numbers. We don’t yet have outcome numbers. Those will matter more.