Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Defense

Pentagon opens AI agent building to three million staff

The Defense Department has added a tool called Agent Designer to GenAI.mil, its enterprise generative AI platform. DefenseScoop reported on March 10 that the Pentagon chief technology officer’s office announced the tool, which is being integrated with Google’s Gemini capabilities, and said three million department employees, including those with no coding experience, can now build “custom AI assistants to automate tasks and streamline complex workflows.”

Google announced the launch on March 9, according to Business Today, which reports that the agents will handle only unclassified tasks.

/ 01What was announced

Per the department’s announcement as reported by DefenseScoop, agents built with the tool “can perform multi-step tasks, ingest various data sources, and be shared with teams for immediate deployment.” The examples offered were practical: operators generating after-action reports, analysts turning images into memos, comptrollers building apps to analyse financial data. Some pre-built agents are also available.

Business Today describes Agent Designer as a low-code tool in which users describe what they need in plain language, aimed at repetitive, multi-step administrative work such as meeting notes, action items and step-by-step project plans. It quotes Google: “This allows users to train and build agents to increase efficiency and reduce repetitive tasks, based on their unique needs.” The same report says GenAI.mil has more than one million unique users and that Google plans regular training sessions with the Chief Digital and Artificial Intelligence Office.

GenAI.mil has been the preferred generative AI platform for all the military services since its launch in December. DefenseScoop notes the department plans to add tools from OpenAI and xAI, and that the announcement did not say whether Agent Designer will connect to them.

/ 02Why it matters

Most organisations treat agent building as something a central team does. This puts it in the hands of the whole workforce. That is a sensible way to find the high-value uses quickly: the person who writes after-action reports knows better than anyone which parts of the job are repetitive. It is also a fast way to end up with a very large number of agents, each with its own instructions, its own data sources and its own idea of what a correct output looks like.

For administrative work on unclassified data, that is a manageable risk. But the examples already include analysis — images into memos, financial data into apps — and the boundary between administrative and analytic work is not sharp. As agents get shared across teams “for immediate deployment,” the questions shift from “does it save time?” to “who checked it, what data does it read, and can we see why it produced this?”

/ 03What agents at this scale need

Three things become important once agents multiply.

Shared definitions. If one agent’s after-action report and another’s use different meanings for the same unit, location or event, their outputs cannot be combined. A common ontology — an agreed model of the things the organisation talks about and how they relate — lets agents built by different people produce outputs that fit together.

Provenance. An agent that “ingests various data sources” should be able to show which source each statement in its output came from. Otherwise a memo produced by an agent is only as trustworthy as the least careful prompt that built it, and nobody downstream can tell.

Repeatability. A report that comes out differently each time the same inputs are run is hard to review and harder to audit. Agents that reason over a structured graph of facts, rather than re-reading raw text each time, can give the same answer to the same question and show the same evidence.

Letting everyone build agents is the easy part. Making their outputs consistent, sourced and repeatable is the work.

/ 04What to watch

  • Agent inventory and review. Whether the department sets up a way to catalogue, review and retire shared agents as their number grows.
  • Model integration. Whether Agent Designer becomes model-agnostic as ChatGPT and Grok arrive on GenAI.mil.
  • Movement up the classification ladder. The tool is for unclassified tasks today. Any move to classified networks will raise the bar for provenance and audit considerably.
How it unfolded
DEC 2025GenAI.mil launches[1]MAR 9, 2026Google announces AgentDesigner[2]MAR 10, 2026Pentagon CTO officeannounces tool[1]MAR 12, 2026This postDEC 2025GenAI.mil launches [1]MAR 9, 2026Google announces Agent Designer [2]MAR 10, 2026Pentagon CTO office announces tool [1]MAR 12, 2026This post
Aktiver’s take — Defense & intelligence

Workforce-wide agent building will surface good uses quickly and multiply inconsistent ones just as fast. Millions of potential authors need one shared vocabulary for units, events and locations.

Treat each agent as an entity in the graph: its author, the data sources it reads, the ontology terms it emits, and the review that cleared it for sharing. An after-action report produced by one team’s agent then uses the same event and unit identifiers as an analytic memo from another, so the two can be joined rather than reconciled by hand. When a source turns out to be wrong, a traversal lists every agent and every product that consumed it. Knowledge Builder works this way, with semantic enforcement over the graph so agents fail loudly when an output breaks the model, and Agent Catalog agents can be adapted to the same ontology. That is what makes an inventory of shared agents auditable.

How we’d model it

AgentStaff memberData sourceAfter-action reportEventReviewauthoredByreadsgeneratedBydescribesclearedByAgentStaff memberData sourceAfter-actionreportEventReviewauthoredByreadsgeneratedBydescribesclearedBy
  • Agent —authoredBy→ Staff member
  • Agent —reads→ Data source
  • After-action report —generatedBy→ Agent
  • After-action report —describes→ Event
  • Agent —clearedBy→ Review

See where this applies →

Sources — read the originals
  1. Pentagon says employees can create their own 'custom AI assistants' with new tech — DefenseScoop defensescoop.com ↗
  2. Google launches custom AI agent builder on GenAI.mil for U.S. military — Business Today businesstoday.in ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading