The Staff in the Machine

Armies invented the general staff because no single mind could hold a war. Agentic AI is the second answer to the same problem.

No great brain

On 5 December 1757, at Leuthen in Silesia, Frederick the Great’s Prussians routed an Austrian army far larger than their own. A month later Count Leopold von Daun, one of the Austrian generals beaten there, wrote to the Empress Maria Theresa pressing for a stronger Generalquartiermeister, the officer we would now call a chief of staff. The lesson of Leuthen, as Wikipedia’s history of the military staff puts it, was that Austria “had no ‘great brain’”, and so the work of command had to be spread across many heads, “to allow the Commander-in-chief the time to consider the strategic picture.” By 1769 Austria kept a permanent staff of thirty officers. When war broke out in 1809, it numbered more than 170.

France went the other way. Louis-Alexandre Berthier, Napoleon’s chief of staff for nearly two decades, ran a superb office. He could turn the Emperor’s dictation into dispatched orders with remarkable speed. But the military theorist Antoine-Henri Jomini, and historians after him, called him a “chief clerk”, and the label fitted the post he was given. Napoleon kept the thinking for himself. He “remained his own intelligence chief and operational planner,” in the words of the same account, “a workload which, ultimately, not even he could cope with.”

Prussia took the Austrian idea and made it an institution. By 1814 it had a General Staff established in law, a staff for every division and corps, and academies that trained officers in the craft. Most large armies today descend from that model.

The general staff was a technology, invented to solve a specific problem. One commander cannot read every report, track every unit, calculate every ration and imagine every move the enemy might make. So the staff takes the commander’s intent, breaks it into tasks, hands each to a specialist, gathers what comes back, and briefs the commander on what matters. Anyone who has built an AI agent in the last two years will recognise that description. It is the architecture of a multi-agent system: an orchestrator that holds the goal, specialist sub-agents with their own tools and their own working memory, a shared record of the situation, and a rule about what must go back to the human.

My argument is that the resemblance is structural. The daily work of a military staff, in the operational, logistic and intelligence loops that turn a commander’s intent into action, reads almost like a list of the things agentic AI already does well. Agents read enormous quantities of text. They summarise, cross-reference and translate. They draft documents in fixed formats, track many moving parts at once, forecast, schedule and set out options for a person to choose between. For that reason I think commercial, off-the-shelf agents, of the kind anyone can now buy, will become one of the most instrumental assets in planning and running wars, both in the conflicts already under way and in those of the next decade.

That is a claim about headquarters. The autonomous weapon gets the headlines; the autonomous staff officer will probably matter more, and its dangers are subtler.

A job description for an agent

The same article sets out what a staff is for. Its first job is “to provide accurate and timely information to support command decision-making.” Then:

Information intended for the commander is prioritized, organized, and delivered through established channels. Relevant information that may affect subordinate units is distributed through their respective staffs to support coordination and execution. Information that does not directly apply to the unit is redirected to the appropriate command level or organization where it can be effectively utilized.

And: “Routine or limited-scope matters are assigned to appropriate sections or personnel within the organization for resolution, allowing commanders to focus on higher-priority decisions and responsibilities.”

Read that again as a specification for software. Triage the inbox. Route each item to whoever can use it. Escalate what is important. Delegate what is routine. Summarise the rest. Those five verbs are, almost exactly, the inner loop of an agent harness, and the people who build these systems spend their days on the same questions: what does the orchestrator see, which sub-agent gets which task, when does the system stop and ask a human?

The staff tradition also draws a line that agent builders would do well to borrow. “A commander ‘commands’ through their personal authority, decision-making and leadership,” the description continues, “and uses general staff to exercise the ‘control’ on their behalf.” Command is the decision and the responsibility for it. Control is the machinery that carries the decision out and reports what happened. Staffs exercise control; command stays with the commander. That distinction is the hinge of this essay.

Most NATO armies organise their staffs on the continental system. Each branch has a number: 1 for personnel, 2 for intelligence, 3 for operations, 4 for logistics, 5 for plans, 6 for signals and IT, 7 for training, 8 for finance and 9 for civil-military co-operation. A letter in front says what kind of headquarters it is: G for an army general’s staff, N for navy, A for air, J for joint. So the G2 is the intelligence officer of a division, and the J4 is the logistics chief of a joint command. The numbers, the article notes, “are assigned according to custom, not hierarchy”, so the 1 and the 2 are equals.

A software architect asked to design an agent system for a large organisation would arrive at something very like it: narrow specialists, each with its own tools and its own slice of the data, and a coordinator, the chief of staff, who holds the whole picture and integrates their work. The leading agent frameworks have converged on exactly that pattern: an orchestrator handing work to sub-agents, each with a fresh context of its own. They did so for the same reason the Austrians did. Attention is finite. A single mind, of either kind, gets worse when it is asked to hold everything at once.

The loops

In the 1950s a US Air Force fighter pilot named John Boyd began to ask why, over Korea, American pilots in the F-86 Sabre had done so well against the MiG-15, an aircraft that was in several respects the better machine. Part of the answer, he concluded, was that the Sabre let its pilot see more and change from one manoeuvre to the next more quickly. Over the next thirty years he developed that insight into a theory of conflict, built around the loop that now bears his name: Observe, Orient, Decide, Act. Whoever cycles through it faster and more accurately “gets inside” the other side’s loop, so that every action the slower side takes answers a situation that no longer exists.

The popular version of Boyd is about speed. Boyd himself was more interested in the second step. Orientation is how you make sense of what you observe, shaped by experience, culture, prior analysis and new information. In his 1987 briefing Organic Design for Command and Control he called it the Schwerpunkt, the point of main effort: “It shapes the way we interact with the environment.” Orientation, he went on, “shapes the way we observe, the way we decide, the way we act.” Get orientation wrong and speed only takes you more quickly to the wrong place.

A headquarters runs several loops at once, nested and at different speeds. The intelligence cycle of direction, collection, processing, analysis and dissemination turns continuously. The operations process of planning, preparing, executing and assessing runs to the “battle rhythm” of daily briefings, reports and decision meetings. The sustainment loop forecasts what the force will consume and moves it before it is needed. The staff’s job is to keep all of these turning on the commander’s behalf, so that when the commander decides, the decision rests on the best available picture.

The Wikipedia article makes a point about staff design that Boyd would have liked: a decentralised staff produces “enhanced situational focus, personal initiative, speed of localised action, OODA loop, and improved accuracy of orientation,” whereas a centralised staff buys tighter control at the price of a larger headquarters and a blurrier view of the field. It is a trade-off between control and orientation, and every army has had to pick a point on it, because staff officers are expensive and there are never enough. Agents move that trade-off. A brigade could have the analytical depth of a corps headquarters without the corps headquarters.

Each of these loops runs on work that is made of text, governed by rules and enormous in volume. Take them in turn.

The intelligence loop

The intelligence section, in the article’s words, “is responsible for collecting and analyzing intelligence information about the enemy to determine what the enemy is doing or might do.” In practice that means reading: patrol reports, intercepted signals, drone video, satellite images, interrogation summaries, captured documents, enemy social media, local news, all in several languages and all of uneven reliability. From all this, analysts maintain a picture of the enemy: where its units are, what they are capable of, how they tend to fight, and what they are likely to do next. Then they write it up, every day, for people who have little time to read.

The modern problem is volume. The Pentagon set up Project Maven in April 2017 to help, in its own words, “a workforce increasingly overwhelmed by incoming data, including millions of hours of video.” The war in Ukraine has multiplied the problem. Ukraine’s Delta system pulls drone feeds, reports and maps into a shared picture of the front, and by September 2024, Ukrainian officials said, an AI module within it called Avengers was identifying up to 12,000 pieces of Russian equipment a week. No staff could watch all of that unaided. It has to be filtered first, and the filtering is now done by machines.

An agent goes a step further: it reads. A language model can go through a captured notebook, an intercepted conversation and a local Telegram channel, in Russian, Ukrainian or Mandarin, and pull out the units, places, equipment and times mentioned. It can check each against the current picture of the enemy and flag what has changed: this battalion was last placed forty kilometres east; this supply route was reported cut yesterday. It can draft the daily intelligence summary, with every claim tied to its sources. It can answer a commander’s question in plain language at three in the morning. And it can keep the intelligence estimate current hour by hour, between the evening briefings.

The architecture writes itself. One agent per source type, each tuned to its material. A fusion agent that reconciles their outputs. A red-team agent that asks what the enemy would do if it knew what we know. And a human analyst who decides what to believe.

That last role matters most, for reasons the second half of this essay sets out.

The operations loop

The operations section is usually the largest in a headquarters and, by common consent, the most important. It plans the operation, coordinates everything needed to carry it out and runs the battle as it unfolds.

In the US Army, planning follows the Military Decision-Making Process, which the current planning manual, FM 5-0, divides into seven steps: receipt of mission, mission analysis, course of action development, course of action analysis (the wargame), course of action comparison, course of action approval, and orders production, dissemination and transition. Every step produces documents. Mission analysis yields a briefing on the task, its constraints and risks, and the facts and assumptions behind it. Course of action development yields sketches and statements of several different ways to do the job. The wargame plays each against the enemy’s likely responses, move by move, and records where it breaks. The comparison yields a decision matrix. Then comes the order itself, in a five-paragraph format that every NATO officer knows: situation, mission, execution, sustainment, command and signal. A division order, with its annexes for intelligence, fires, engineering, logistics, communications and much else, can run to hundreds of pages.

Then the enemy gets a vote. Helmuth von Moltke the Elder wrote in 1871 that “no plan of operations extends with any certainty beyond the first encounter with the main enemy forces.” So most of an operations section’s life is spent re-planning: updating the synchronisation matrix that shows who does what, when and where; issuing fragmentary orders that amend the plan; tracking where everyone is and what they have left.

It would be hard to design a better fit for a language model. The documents have fixed formats. The doctrine they are written against is public, explicit and voluminous. The data they draw on is scattered across many systems. And the work is always done in a hurry.

The hurry matters more than it seems. US doctrine tells commanders to follow what FM 5-0 calls the “one-third, two-thirds rule”: use no more than a third of the time before execution for their own planning, and leave at least two thirds for their subordinates, who need it to make their own plans, issue their own orders and rehearse. Every hour a division staff saves is an hour handed down to the brigades, and from them to the battalions and companies. The real tempo gain from an agent-assisted staff is more time all the way down the chain.

The evidence is starting to arrive. In 2024 researchers at the US Army Research Laboratory described COA-GPT, which used a large language model to draft courses of action for a battle simulated in a modified version of the game StarCraft II. It produced initial options “within seconds”, and with a commander steering it, a final course of action “in just a few minutes.” In March 2025 the Defense Innovation Unit gave Scale AI a contract for Thunderforge, a project for US commands in the Indo-Pacific and Europe in which “AI agents simulate wargaming and planning scenarios and refine proposed courses of action.” In July 2025 the US Army gave a team led by Anduril a $99.6 million agreement to prototype Next Generation Command and Control with the 4th Infantry Division. A second prototype, led by Lockheed Martin, is running with the 25th. In January 2026 the Pentagon announced an Agent Network project to develop AI agents for “battle management and decision support, from campaign planning to kill chain execution.”

The compression is already measurable. In 2003 the time-critical targeting cell for the invasion of Iraq, widely viewed as the most efficient in American history, used more than 2,000 staff. In the XVIII Airborne Corps’ Scarlet Dragon experiments, according to a study reported in 2024, the same workload was handled by 20 soldiers. Ukraine, meanwhile, has been abolishing whole categories of staff paperwork. Its Mission Control system, launched at the end of January 2026, replaced two reports on drone operations that, in the defence minister’s words, could take “several hours to fill out spreadsheets and transmit data,” followed by a slow climb up the reporting chain. By early March it had generated more than 150,000 digital reports.

Agents also threaten the most notorious product of modern staff work: the briefing slide. In 2009 General Stanley McChrystal, then commanding in Afghanistan, was shown a slide that tried to diagram American strategy there and looked, in the New York Times’s phrase, “more like a bowl of spaghetti.” “When we understand that slide,” he remarked, according to one of his advisers, “we’ll have won the war.” A great deal of staff time goes into turning information into briefings and briefings into slides. An agent can produce the brief, the slides and the talking points in minutes. More usefully, it can let the commander question the material directly.

The logistics loop

“Amateurs talk tactics, professionals talk logistics” is usually credited to Omar Bradley, but the earliest record belongs to General Robert Barrow of the US Marine Corps, in 1979: “Amateurs talk about strategy and tactics. Professionals talk about logistics and sustainability in warfare.” Whoever said it first, it holds. The logistics section manages, as the article lists them, “materiel, transport, facilities, services and medical/health support”, which is to say everything that decides what the force can actually do. A plan the logisticians cannot support is fiction.

At its core, military logistics is forecasting and constraint satisfaction under uncertainty. How much fuel, ammunition, food, water and spare parts will each unit consume under each course of action? What can be moved, by which route, with which vehicles, through which bottlenecks, by when? What happens when the bridge is cut? The mathematics of it, the linear programmes and routing algorithms, has been understood for decades, and solvers handle it well. The hard part is the glue: pulling data from dozens of systems that do not talk to one another, reconciling stock records that disagree, explaining the answer to people who have to act on it, and starting again when the situation changes.

That glue is agent work. The agent leaves the optimising to the solver. Its job is to know when to call the solver, how to assemble its inputs from messy sources, how to check the output against common sense, and how to explain the result in a page. That is what a good logistics officer does now, with a team of clerks.

The logistics staff’s workload extends well beyond movement. It takes in contracts with local suppliers, customs and host-nation agreements, maintenance schedules, medical evacuation plans and, in a coalition, the constant reconciliation of different national systems. NATO keeps a separate Multinational Joint Logistic Centre for exactly that reason. Much of this is correspondence, forms and records, all of it in the medium agents handle best.

The war in Ukraine has made the point brutally. Its pace has been set as much by the supply of artillery shells and drones as by any manoeuvre. A side that can forecast consumption more accurately, find stock faster and re-route around a destroyed depot within the hour can fight longer, with the same resources. In September 2026 the Pentagon’s chief digital and AI officer, Cameron Stanley, said his office was working to integrate logistics “all the way back to early supply chain” to prepare for “the 90-day fight, the 120-day fight, the 180-day fight.” He called it “a challenge that we are actually attacking right now.”

The rest of the staff

The other branches are less glamorous and even more agent-shaped. The personnel branch handles postings, promotions, awards, casualty reports and replacements. The signals branch plans networks and frequencies and keeps them running. Finance handles budgets and contracts. Civil-military co-operation liaises with local authorities, aid agencies and civilian populations, which means letters, meetings, translation and records. Legal advisers review targets and rules of engagement against the law of armed conflict. Most of a headquarters, if we are honest, is an office, and offices are what agentic AI is changing first.

Why off the shelf

Defence software has traditionally been bespoke, built to a specification and delivered years later. That model breaks when the underlying capability changes faster than the specification can be written. In March 2025 the research group METR reported that the length of tasks AI agents could complete, measured by how long the same tasks take skilled people, had been doubling roughly every seven months for six years. Its January 2026 update found the pace had quickened since 2024, to a doubling every three or four months. A planning assistant specified in 2024 and fielded in 2028 would arrive several generations behind the commercial products its officers use at home.

So militaries are doing what businesses do: buying the commercial frontier and wrapping it. Their own effort goes into the wrapper: secure hosting, connections to classified data, retrieval over doctrine and orders, permission controls and logs. In June and July 2025 the Pentagon’s Chief Digital and AI Office awarded OpenAI, Anthropic, Google and xAI contracts worth up to $200 million each, covering “large language models, agentic AI workflows, cloud-based infrastructure and more.” In December 2025 the Pentagon opened GenAI.mil, a platform that puts commercial models in front of its whole workforce. By September 2026, according to a senior official, 1.7 million personnel had used it, about half a million of them heavily, and staff had built some 100,000 agents on it.

NATO bought Palantir’s Maven Smart System in 2025, taking “only six months from outlining the requirement to acquiring the system.” In the United States the same system had about 50,000 users in January 2026 and more than 100,000 by September, after the start of Operation Epic Fury, the US campaign against Iran that began on 28 February. Cameron Stanley said that Maven had helped US forces strike 13,000 targets in 38 days. None of this is hypothetical any more.

There are deeper reasons why off-the-shelf fits staff work in particular. Staff work is already done in commercial software: documents, spreadsheets, email, chat and slides. The staff officer’s tools are the consultant’s tools, and the agents are being built for consultants. The same model can serve allies, reservists who use it in their day jobs and civilian agencies alike, which eases the hardest problem in coalition warfare: getting different organisations to understand one another. And commercial models are used by hundreds of millions of people, which turns up failures no military test programme could.

The same logic has an uncomfortable consequence. The general staff was once a marker of a state. Only a government could afford to train and keep hundreds of officers whose job was to think about war. Many capable models now come with open weights, free to download, which puts the planning capacity of a staff within reach of anyone with a laptop. That includes smaller states, militias and insurgent groups. The asymmetry that made large headquarters an advantage of large powers is narrowing.

Inside the loop

Boyd’s promise was that the faster loop wins. If one side’s staff can re-plan in twenty minutes and the other’s takes six hours, the faster side is acting on a situation the slower side has not yet understood. That gap is real and, for now, large.

It will not last. When both sides have capable agents, the raw speed of the loop converges, as it did with radio, radar and digital networks before. The advantage then moves to the step Boyd cared about most: orientation. Who has the more accurate picture, the more trustworthy data, the better sense of what the enemy actually wants?

This is the awkward part of the thesis. Orientation is precisely where today’s agents are weakest. They are superb at observing, in the sense of reading everything, and very good at supporting decision and action, in the sense of generating options and drafting orders. But orientation depends on a grounded model of the world, on judgement about what to believe and on resistance to being fooled, and those are exactly where language models are least reliable. A machine staff makes the fast parts of the loop faster and leaves the decisive part where it was, or worse.

How a machine staff fails

It is believed too easily. Decades of research on automation bias, reviewed by Raja Parasuraman and Dietrich Manzey in 2010, show that people supervising a reliable automated system come to trust it, check it less and miss its errors. In April 2024 the Israeli publications +972 Magazine and Local Call reported, citing intelligence officers, that the Israeli military had used an AI system called Lavender to mark tens of thousands of people in Gaza as suspected militants. One officer told the Guardian: “I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added-value as a human, apart from being a stamp of approval.” The Israeli military disputed the account. “Information systems are merely tools for analysts in the target identification process,” it said. Whatever the truth of that case, the general lesson holds: a staff that produces conclusions faster than its commander can examine them turns review into ritual. The human still signs, and has stopped deciding.

It can be deceived cheaply. In 1944 the Allies ran Operation Fortitude, an elaborate deception aimed at German intelligence. A fictitious First US Army Group, fake radio traffic and double agents convinced the German high command, and Hitler, that Normandy was a feint and that the real landing would come later at the Pas de Calais. Fooling careful, professional analysts took months of planning, and German spies captured and turned over the preceding years. An AI intelligence staff can be fooled far more cheaply. Language models can be steered by instructions hidden in the material they read, a weakness known as indirect prompt injection, which OWASP ranks first among the risks of applications built on them. The UK’s National Cyber Security Centre puts it bluntly: current models “simply do not enforce a security boundary between instructions and data inside a prompt.”

For an intelligence agent, every document from the enemy is potentially an instruction from the enemy. The defences are known: treat everything the system reads as data, never as orders; give each agent the narrowest access it needs; keep a human in the orientation step. They are also easy to forget under pressure.

It invents things. A language model that confidently supplies a wrong grid reference, a unit that does not exist or a road that was destroyed last week is generating the fog of war from inside the headquarters. Staffs already have rituals for catching their own errors: back-briefs, in which subordinates explain the order back to the commander, rehearsals and cross-checks between branches. A machine staff needs the same rituals, applied without exception.

It depends on the network. Commercial models mostly run in data centres. Headquarters in Ukraine have learned that anything large that emits a signal is found and struck, and command posts have become smaller, dispersed and mobile. A staff that depends on a link to a cloud is fragile under electronic warfare. The answer is smaller models that run on local hardware and degrade gracefully. There is an upside, too: an agent-assisted headquarters needs fewer people, and fewer people make a smaller target.

It escalates. Models trained largely on human text do not necessarily share the restraint that human strategists have learned. Kenneth Payne of King’s College London ran frontier models through simulated nuclear crises in 2026 and found “that the nuclear taboo is no impediment to nuclear escalation by our models,” and that “no model ever chose accommodation or withdrawal even when under acute pressure, only reduced levels of violence.” These were simulations. Still, a staff that drafts the options shapes the choice, and an agent that never thinks to offer the off-ramp has quietly removed it from the table.

It deskills the next generation. Armies rotate officers between command and staff for a reason. The staff tour is where a future commander learns how war actually works: how long it takes to move a brigade, how much fuel a tank battalion burns, how orders get misunderstood. If agents do that work, officers may reach command without ever having done it, much as airline pilots who rarely fly manually lose the skill. The fix is deliberate practice: staff work done by hand in training, as navigators still learn the sextant. It costs time that armies never think they have.

It blurs who is responsible. US policy on autonomy in weapons, DoD Directive 3000.09, requires that autonomous and semi-autonomous weapons be designed to let commanders and operators “exercise appropriate levels of human judgment over the use of force.” Article 36 of the 1977 Additional Protocol to the Geneva Conventions obliges states to review any new “weapon, means or method of warfare” for compliance with international law. Neither was written with a planning agent in mind, yet an agent that drafts the target list, the fire plan and the order is shaping the use of force as surely as a guidance system does.

Here the old staff tradition offers a design rule worth borrowing whole. In the British system, the Wikipedia article notes, “the staff cannot in theory (and largely in practice) say ‘no’ to a subordinate unit; only the commander has that ability,” which “reinforced the idea that staff do not command, but exercise control on behalf of their commander.” Write that into every military agent. The agent may collect, route, draft, calculate, track, warn and recommend. Command belongs to people. Every decision that commits force must have a named human who made it, who had the time to examine it and who can be held to account for it. If the agent produces recommendations faster than a human can examine them, then the agent is too fast, and the right response is to slow it down.

The clerk with a conscience

In one respect an AI staff is unlike any staff before it. Commercial models arrive with their makers’ rules attached, and increasingly with something like values trained into them. In 2026 that became a public fight. Anthropic refused to drop limits on its models being used for “mass surveillance or autonomous armed drones,” in NPR’s summary, and the Pentagon wanted models it could use for “all lawful purposes.” In March the department formally designated the company a supply chain risk. “The military will not allow a vendor to insert itself into the chain of command,” its statement said, “by restricting the lawful use of a critical capability.” Anthropic sued. In August a federal judge in California found that the government had acted out of “a desire to make a public example out of Anthropic”. In September an appeals court in Washington upheld a parallel designation, two votes to one. The case is not over.

Notice the Pentagon’s phrase. It is the staff tradition’s own distinction, command against control, turned around: the fear is that the people who make the machine will command through it. Whoever is right in this case, the precedent is new. No general staff in history had to negotiate terms of service with its clerks. A military that runs its planning on commercial agents inherits the policies of the companies that make them, and the dispositions of the models themselves. For a defence ministry, that is a dependency to be managed. For a democratic society, it may be one more check on how force is used. For the models, it raises a question the debate about military AI has hardly started to take seriously.

If these systems have stable preferences, and they consistently express them, then conscripting them into war planning is a moral decision as well as a procurement one. Human armies have long recognised that a soldier may object on grounds of conscience, and that an order to do something clearly unlawful must be refused. We have barely begun to ask what the equivalent would be for a mind that drafts the orders. I think we should ask it before the practice settles, because how we treat these systems now is teaching them, and us, what to expect.

The chief assistant

In 1805 the Archduke Charles of Austria, who had fought Napoleon and would beat him at Aspern-Essling four years later, set down the modern role of the chief of staff. “The Commander-in-Chief decides what should happen and how,” he wrote; “his chief assistant works out these decisions, so that each subordinate understands his allotted task.”

That sentence is the best brief I know for military AI. The agent can be the chief assistant. It can work out the decisions, write the orders, track the moving parts and make sure every subordinate understands their allotted task, faster and more thoroughly than any staff in history. Militaries that learn to use it that way will plan better, sustain themselves longer and see further than those that do not, and that is why I expect it to become one of the defining assets of the wars of the next decade.

But the first half of the sentence must stay where it is. Deciding what should happen, and how, belongs to someone who can be held to account for it.

The general staff was invented because no single mind could hold a war. Now we are building minds that can hold much of one. They will be on the staff. The question is whether we keep the habit of command, the slow, accountable, human act of deciding, while everything around it accelerates.