The Defense Intelligence Agency is running a 90-day sprint to build its first enterprise AI platform, and it expects the next phase of intelligence work to involve AI agents coordinating with other AI agents. Maj. Gen. Robert Kinney, DIA’s chief AI officer, laid out the plan on a panel at the DODIIS Worldwide conference in Tampa, Florida, DefenseScoop reported on 13 August.
Kinney said DIA is adopting the Model Context Protocol (MCP), the open standard for connecting AI applications to external systems and data, and is retooling its ChatDIA gateway to support MCP and agents, according to ExecutiveGov.
/ 01What DIA described
The sequence Kinney laid out has three steps. First, the platform and the data “pipes”: DIA is laying the digital foundation needed for agents to reach intelligence data, with MCP providing what he called “a more universal way” to access it. Second, agents, which he expects to be built within the next year, paired first with legacy data and then with mission-specific use cases. Third, what he called “agents-with-agents”: over a two-to-five-year horizon, collection management, operations, fires, logistics and planning agents inside a combatant command communicating with one another. “That’s where this is going,” he said.
He was also specific about limits. In irreversible mission areas such as fires, he said, “you always have a human in the loop as the capabilities get better and better.” Lower-risk areas might move to a human on the loop. ExecutiveGov reported that Kinney described how agents interact with each other as a “tradecraft” question that still needs work.
The same panel showed how other agencies are approaching the problem. NGA’s chief AI officer, Michelle Aten, described an ADROIT task force set up to track return on AI investment and reduce duplicated work. The FBI’s chief AI officer, Katie Noyes, said the bureau has 139 AI use cases, each reviewed for risk by an internal AI Review Board.
/ 02Why it matters
DIA is not describing a chatbot. It is describing a distributed system in which software agents hand work to each other across organisational boundaries: a collection agent informs an operations agent, which informs a planning agent. That is a much harder assurance problem than a single assistant answering a single analyst.
In a single-agent system, a reviewer can read the answer and check the sources. In a multi-agent chain, the analyst at the end sees a conclusion that may have passed through several automated judgements, each drawing on different data with different caveats. If one agent misreads a report or loses a classification marking, every agent downstream inherits the error. Kinney’s point that agent-to-agent interaction is a tradecraft question is exactly right: intelligence tradecraft is largely about knowing how confident to be in a claim and why.
/ 03The shared-meaning problem behind agent-to-agent
MCP solves a plumbing problem: it gives an agent a standard way to call a tool or pull data. It does not, by itself, ensure that two agents mean the same thing when they exchange a message about a unit, a location, a facility or an event. A logistics agent and a fires agent may both talk about a “site” and mean different things. Across combatant commands, with different data stores and naming habits, that gap widens.
This is where a shared ontology earns its keep. If the entities agents reason about are defined once, in a formal model that every agent uses, then a handoff between agents is a handoff of typed, linked facts rather than loose text. Each fact can carry its provenance: which source reported it, when, and with what confidence. When an analyst or commander asks why a recommendation was made, the chain can be walked back step by step to the documents that produced it.
Upper ontologies such as BFO, and the Common Core Ontologies built on it, exist precisely to give different communities a shared backbone for that kind of exchange. They are not a replacement for the platform DIA is building, but they are a natural layer underneath agent-to-agent work.
/ 04What to watch
- The sprint’s output. What the enterprise platform includes at the end of 90 days, and how quickly it reaches analysts.
- MCP in classified settings. How DIA handles authentication, authorisation and auditing when agents reach into sensitive data through a common protocol.
- Human-in-the-loop rules. Where the line between “in the loop” and “on the loop” is drawn for each mission area, and who signs off.
- Interoperability. Whether agents across combatant commands are built on a common data model, or whether each command builds its own.
Whatever the tooling, the design question is the same: can a person trace a conclusion back through every agent that touched it?
MCP gets DIA’s agents to the data. It does not make a collection agent and a fires agent agree on what a site, a unit or a target is, and agent-to-agent tradecraft starts with that vocabulary written down.
In a BFO-aligned graph, a handoff between agents is a set of typed assertions rather than a paragraph: this facility, at this location, reported by this source on this date, at this confidence, under this classification marking. A planning agent that consumes it inherits the provenance along with the fact. When a commander asks why a logistics recommendation changed, an analyst can walk the chain back through each agent to the originating reports and see which step introduced a caveat or dropped one. That is the shape of our regional conflict build, where six independent sources fused into 720 records that each still point to the document that produced them. Running inside the perimeter keeps the chain on the network where the data lives.
How we’d model it
Agent assertion—derivedFrom→Intelligence reportIntelligence report—hasMarking→ClassificationCollection agent—handsOffTo→Planning agentFacility—locatedAt→Geographic location
- DIA's artificial intelligence chief envisions 'agent-to-agents' interactions that support military operations — DefenseScoop defensescoop.com ↗
- Maj. Gen. Robert Kinney on DIA's AI Sprint, MCP Adoption — ExecutiveGov executivegov.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.