Agentic AI and the Future of the U.S. Military
A Roundtable of the Special Competitive Studies Project
SCSP Headquarters · Arlington, Virginia · Summer 2026
The character of war is changing — not incrementally, but fundamentally. SCSP convened more than two dozen former senior commanders, defense officials, frontier-AI technologists, and practitioners with direct experience of the wars in Ukraine and the Middle East to answer one question: Does the U.S. military need radical transformation, or significant reform?
Read moreOverview
A Hinge Moment
On July 21, 2026, the Special Competitive Studies Project convened a private roundtable on Agentic AI and the Future of the U.S. Military. General CQ Brown, Jr., 21st Chairman of the Joint Chiefs of Staff, and General Joe Dunford, 19th Chairman, co-chaired the session; General David H. Petraeus, who commanded the Surge in Iraq, U.S. Central Command, and coalition forces in Afghanistan, and subsequently served as Director of the CIA, framed the discussion.
Participants included more than two dozen former senior commanders and defense officials — among them former combatant commanders, military service chiefs, intelligence agency directors, and allied defense chiefs — together with leading technologists from America's frontier AI companies and practitioners with direct experience of the wars in Ukraine and the Middle East. Collectively, the room represented more than 900 years of service to national, allied, and international security.
The convening came at a hinge moment. The character of war is changing — not incrementally, but fundamentally — and artificial intelligence is advancing from reactive, generative systems that produce text and content to proactive agentic systems that pursue goals, perceive their environment, plan across multiple steps, act in the real world, and learn continuously from what they do. Militaries that grasp and adapt to this shift will fight differently. Those that do not will discover the difference the hard way, as militaries throughout history always have.
The roundtable was held under the Chatham House Rule; this report — prepared by SCSP staff — summarizes the substance of what was said without attributing remarks to individual participants or their organizations.
The Central Question: Radical Transformation or Significant Reform?
The discussions at the roundtable centered on a single strategic choice: given the advances in AI and adjacent technologies — above all, unmanned and increasingly autonomous systems — does the U.S. military need to radically transform how it fights wars and projects power, and undertake correspondingly ambitious institutional transformation? Or will significant reforms in certain areas prove responsive to the changes now underway in the character of war and the advances in technology? The weight of the discussion that followed fell clearly on the transformation side of the ledger: as one participant put it, the question is not whether revolutionary change is needed but how to achieve it. There was consensus that throughout military history, success has rarely gone to those who merely acquired a new technology; it has gone to those who adopted it and adapted concepts, doctrine, organizational structures, training, procurement, and leader development to exploit that technology most effectively. The tank, the aircraft carrier, and precision strike each taught the same lesson. The advent of massive numbers of inexpensive, remotely piloted systems, the looming introduction of truly autonomous unmanned systems, and the incorporation of Agentic AI will underscore that lesson again.
By this standard, the United States is falling short at the moment. The U.S. military has not embarked on the truly institutional change required by what is visible on the battlefield in Ukraine, on the Black Sea, and in the skies over Crimea and the Russian Federation, or in the war with Iran and Israel's fight with Hezbollah. Nor has any allied nation in Europe, the Middle East, or the Pacific. China, though, which possesses both the manufacturing capacity the future of war demands and the advanced AI models, and produces vastly more unmanned systems and robotics than any other country — may be absorbing the lessons from Ukraine and the Middle East more assiduously than the countries supporting Ukraine and fighting Iranian aggression. They have the manufacturing capability to scale, but may not have adapted concepts, doctrine, organizational structures, training, procurement, and leader development to maximize the capability. Nevertheless, we should not comfort ourselves with the assumption that our adversaries are studying these wars less seriously than we are. The evidence suggests the opposite.
Indeed, there was broad agreement that there is an urgent need for a revolutionary new concept of warfare that should then drive rapid institutional change – in organizational structures, how we train, the education of leaders, what we buy and how we buy it, and even personnel policies and bases and facilities.
The changes necessitated by agentic AI, some of the attendees argued, may be even greater. First, the battlefield itself is already being remade: Ukraine now employs some 10,000 drones per day, unmanned systems inflict about 95 percent of casualties on Russian soldiers along the front, and Ukraine is now reportedly manufacturing ten million drones a year. The next phase of combat in Ukraine will most certainly feature truly autonomous unmanned systems, and beyond that, autonomous systems of autonomous systems — large interconnected networks of intelligent platforms operating collaboratively across every domain. Relatively low-cost systems are already fundamentally challenging the relevance of and imposing extraordinary costs on exquisite platforms, and there is at present no effective area defense against the drone swarms that will be made possible by truly autonomous drones. Second, and potentially more consequential, agentic AI will serve as the connective tissue linking and enabling humans, autonomous systems, and military organizations. For commanders, agentic AI will amount to an essentially unlimited number of exceptional staff officers — planners, analysts, logisticians, trainers — working continuously, at machine speed, and at the level of Nobel laureates rather than merely strong PhDs. In other words, autonomous systems will provide mass, autonomous systems of autonomous systems will employ mass at machine speed, and agentic AI will orchestrate it all.
The Evidence From Two Wars
Delta: From Pilot to Backbone
The IDF Exploitation Bottleneck
Ukraine: The Battlefield Votes Weekly
Despite the technology's power, the most decisive innovations were institutional:
- 2023 legal decision — battlefield data permitted on commercial cloud infrastructure
- Readiness reporting — units graded on how fast they learn
- Doctrine trails practice — rather than gating it
- Strike-for-points procurement — verified strikes earn points units spend directly on drones
Israel: Rebuilding Around Operational Logic
Facing an exploitation bottleneck of staggering scale in a war the force had not planned for, the IDF succeeded through:
- Mission hacking — narrow optimization of models to specific missions
- Empirical testing — on untidy operational data, during combat
- Short feedback loops — among users, developers, and commanders
- Technologist-reservist pipeline — dual-expert talent surged into the fight within days
“The real challenge is not acquiring the technology — but adapting it into existing operational logic, and adjusting organizational culture to fit it.”
Roundtable participant
The Technological Frontier of AI
Linear Extrapolation Is a Recipe for Strategic Surprise
The discussion also highlighted that advances in AI systems are compounding faster than even sophisticated observers appreciate. Linear extrapolation from the present is a recipe for strategic surprise. One expert described agentic workflows compressing software release cycles from six months to, in some cases, multiple releases per day, and noted that as agents proliferate, the bottleneck shifts from processing data to human decision-making — an approvals problem that forces deliberate choices about which decisions to delegate downward to agents. One of the experts noted that published capability assessments generally underestimate the capabilities, because no one has had the time or the compute power to push models to their limits. Three years ago the best AI systems struggled with grade-school arithmetic. Last year they earned gold medals at the world's premier mathematics olympiad. Today they contribute to the frontier of mathematical research. This pace will persist. Plans predicated on the present state of the art will be obsolete before they are executed.
The full capability of AI systems becomes even more apparent when models have more time to work on a task. Models that score 20 percent when forced to answer immediately reach 80 or 90 percent when allowed to reason for hours or days. Nearly absent from public discussion is what such systems could accomplish if set to work for weeks or months against the hardest strategic problems: campaign design, economic competition, adversary red-teaming. Multi-agent systems, in which a high-level goal is decomposed into sub-goals assigned to and monitored across many agents, were described as the most promising path, shifting the human role from planning to defining objectives, constraints, and resources.
Three Years of Capability
Time Is a Force Multiplier
What Must Be Transformed
Eight Areas. Ninety Days Each.
In the strategic discussion that ensued, participants identified eight key areas in which examination and change is most required if the Joint Force is to respond to the realities seen in Ukraine and also to leverage agentic AI and related technologies at the speed the moment demands. Participants also cautioned against framing the effort as starting from zero. Bottom-up adoption is already underway. But fielding a tool is not transforming an institution, and the distance between the two is what the areas below describe. Given the urgency, participants suggested 90-day sprints should be undertaken on each of these areas:
01. Concepts of operation and doctrine
The first area that the participants identified is the need for urgent development of an entirely new concept for warfare that includes how agentic AI will drive new concepts of operations — and how to ensure the military is capable of executing those concepts. Participants agreed that the traditional relationship between doctrine and practice must be inverted for this era: in the twenty-four months it takes to turn a battlefield lesson into a field manual, the war in Ukraine has changed character three times. Ukraine lets doctrine trail practice, with tactics, techniques, and procedures evolving and spreading at the speed they succeed – because they develop concepts and practices that are adequate to drive and accommodate the constant changes. Only later do they capture it all in formal doctrine manuals.
02. Command structures, staffs, and the combatant commands
Multiple participants argued that integrating AI agents at scale will require an overhaul of command and control structures, with traditional staff models becoming obsolete as agentic systems assume much of the work of planning, monitoring, and synchronization. The discussion also pointed to the deeper shift: as autonomous systems proliferate, agentic AI itself becomes the command-and-control architecture, and the human challenge moves from controlling platforms to commanding systems of systems in accordance with commander's intent and the law of armed conflict. Commander's intent will need to be provided at a pace that can stay ahead of the speed of technology and a staff of AI agents. And, of course, all of this will need to be accompanied or preceded by massive and urgent organizational and force structure changes to reflect the new realities of warfare that are captured in the overall new concepts.
Participants further discussed how mission command may need to be redefined for non-human initiative. Current doctrine rests on human-to-human trust, and has no category for delegation to machines. Participants discussed whether doctrine will need to change to extend mission command to agentic systems, with the degree of human control scaled to the system’s technical capability and the consequences of error. It was noted that some functions will require definitive human control no matter the degree of technical capability.
3. Data and compute as weapons systems
Echoing Delta's core lesson that “data is a weapon system,” participants converged on data architecture as a foundational requirement: continuous pipelines feeding operational data back into training environments so models learn at battlefield tempo, managed with the discipline of a weapons program. Ukraine's single most consequential enabler was identified as a policy, not a technology — the reclassification decision that let enemy order-of-battle data sit legally on commercial clouds. A comparable U.S. classification regime that lets threat data reside where it is useful was flagged as among the hardest and most necessary changes, in order to enable the movement of threat information more freely across the Joint Force, and to allies and industry, at the tempo operations demand. Participants extended the same logic to compute. It must be instrumented and deliberately allocated across the Joint Force, with capacity intentionally reserved for bottom-up experimentation rather than consumed entirely by enterprise programs.
4. Testing, evaluation, and certification of non-deterministic systems
Participants identified a major policy hurdle in certifying agentic AI systems for military operations when their behavior cannot be predicted in every scenario. This connects to a threat highlighted from Ukraine: the unsolved problem of evaluating agents against an adversary who poisons data inputs and updates the test set every six to twelve weeks. Participants stressed that as reliance on agentic AI grows, the integrity of data and models becomes a single point of failure, and that vulnerability to deception and adversarial manipulation — a governance risk — must be engineered against from the start, with human judgment retained as indispensable. Participants stressed that moving fast and in an agile manner is necessary but insufficient — unless the policies and procedures governing usage and employment keep pace with the technology, the force will move only at the speed of cognitive decision-making.
5. Talent and personnel systems
Drawing on the Israeli account of elite technical-talent programs and technologist-reservists, participants argued the United States needs a way to recruit, retain, and promote highly technical personnel in uniform — a purpose for which the current personnel system is not designed. The Israeli model of developing people expert in both technology and operations, and surging cleared technical talent into military development roles within days, was held up as the fastest scaling mechanism any force has demonstrated. Participants discussed the need for executive education focused on agentic AI for general and flag officers. Leaders cannot direct a transformation they do not understand.
6. Procurement and the adaptation cycle
The group returned repeatedly to speed. If a software fix or new drone algorithm is needed at the front, the update should take eighteen hours, not eighteen months in an acquisition cycle: active theaters must be treated as combat laboratories with a direct tie to developers. Ukraine's battlefield-driven procurement — where verified strikes earn points that units spend in a state marketplace, so that the battlefield votes weekly on what gets bought — was cited as proof that resourcing can run at battlefield tempo. Participants noted that current procurement and operational models are simply not designed for the required agility, and that faster integration of commercial AI capabilities, secure environments for testing and deployment, and training personnel to operate alongside highly capable agents are all overdue. It was specifically discussed that sequential approval processes hinder change, especially in comparison to one-stop-shop or parallel review processes, which several described as a combat capability rather than an administrative task.
7. Culture and leadership
Perhaps the most consistent theme across both case studies and the discussion was culture. Ukraine protected its “misfits” instead of promoting them out or crushing them — what one recent report has called “responsible rebellion” — and its leadership ratified adaptation it did not order. Participants urged fostering an environment in which failure in testing and development is accepted and learned from rather than punished, and cautioned against “admiring” AI rather than implementing it — while thinking hard about how adversaries might implement it in ways we would not.
8. Allies, partners, and international norms
Participants urged that agentic AI be approached with allies with an “integrated by design” attitude — starting at the beginning with allies and partners rather than leaving them behind. It was suggested that bilateral or mini-lateral arrangements with tech-forward allies such as the United Kingdom, Australia, or Israel may field integrated agentic capabilities faster than broader alliance mechanisms. The group also flagged the normative dimension: the United States must lead the world in defining the legal and ethical standards governing autonomous action in combat, or others will define them instead – and not to our advantage. Relatedly, participants mentioned the unresolved question of vendor dependency: plans to employ agentic AI assume access to frontier commercial models, yet the norms governing their use in lethal operations remain unsettled — including in the government’s own relationships with frontier AI labs — and understanding how that dependency would affect implementation is a necessary task.
Several participants argued for the creation of an open technical architecture and marketplace of ideas — an internal "app store" in which field-built tools compete on merit and deploy without multi-year test cycles — with explicit top-cover for the junior innovators and constructive misfits already experimenting and building.
Can the Current Defense Enterprise Deliver?
The Warfighter Will Change Quickly. The System Will Not.
Some participants voiced pointed skepticism — not about the necessity of the transformation, but about its likelihood under current institutional arrangements. They described a non-virtuous cycle that presently exists at the Pentagon:
The Non-Virtuous Cycle
- Requirements processseeks to drive change
- Military servicesContest and seek to append the new requirements
- Various Pentagon Entitieswant to control aspects of the change
- Systems Commandswant to field new capabilities
- Members of Congresswant to build them in their respective districts
- Companieswant to provide them with IP rights attached
This SystemMust Be Upended
Urgency itself was identified as the missing ingredient. Ukraine transformed rapidly under survival urgency; Israel did so under pressure from a multi-front war; the United States is attempting to do so with very little urgency at all. That urgency, participants argued, will have to be created inside the Department of War and argued before Congress.
History also reinforces skepticism around rapid change. A participant shared four historical examples – rifles, breech-loaded firearms, tanks, and smart bombs – that showed the defense bureaucracy moving slowly to embrace new capabilities.
The resourcing and authorities dimension compounds the challenge. A revolutionary change “resourced at the tempo of the problem” requires reprogramming, divestment, or new appropriations. Several of the practical changes – from a new data-classification regime to procurement-cycle compression to a cross-command open architecture – plausibly require new statutory authority. Congress, in other words, cannot be an afterthought: a deliberate congressional engagement track is as essential to this transformation.
The discussion nonetheless pointed toward how the change could be made to happen. One suggested approach: Educate, Enable, Empower — educate the force about agentic AI and give people the means to learn; enable bottom-up participation through an open backbone architecture that permits “natural shaping” rather than imposed shaping; and empower people to add to the internal marketplace of capabilities described above, which implies loosening the reins and explicitly accepting risk. Recent structural openings offer vehicles for exactly this: the sub-unified command focused on autonomous warfare created in April 2026, and the newly announced Direct Reporting Portfolio Manager for Unmanned Systems reporting directly to the Deputy Secretary. Ukraine’s own institutional response, including the creation of a separate unmanned-systems service in 2024, was flagged as a case study worth examining. The measure of this effort, participants concluded, is not only whether it says what should be done, but whether it offers leadership specific mechanisms for how.
The Institutional Proposal
USPACOM as the Test Bed
One of the immediate institutional proposals of the session argued that the U.S. Pacific Command should become the test bed for the adoption of new technologies — agentic AI foremost among them, together with unmanned and autonomous systems. Given the pressing nature of the Indo-Pacific challenge, USPACOM's operational demands should drive the transformation: new capabilities should be fielded, evaluated, and iterated against that theater's real and prioritized problems rather than diffused evenly and therefore thinly – across the joint force.
Participants framed the strategic imperative in equally direct terms: credibly undermining Beijing’s confidence in its ability to act against U.S. interests — shaping and deterring its choices — is the problem around which to focus the joint force, and precisely the kind of forcing function that can create the conditions for rapid, revolutionary change.
Participants with direct visibility into the command also observed that USPACOM is, in important respects, already leading the way. Among the lessons drawn from the Command’s use of AI to date: transformation depends on significant input and direct leadership from the commander — the commander, not the staff, owns the risk — making culture the number-one opportunity and the number-one risk; staff engagement up front and throughout implementation is essential to overcome fears that “AI is taking my job”; headquarters-level cross-functional coordination and prioritization of AI efforts is indispensable. The gaps are equally instructive: many AI tools and operational plans still sit on different networks; there is as yet no systemic training and education for AI; and implementation requires clear rules of the road that are consistent across the organization.
To further drive and resource the changes, it was recommended that the Deputy Secretary of War, joined by the Vice Chairman of the Joint Chiefs of Staff, hold weekly meetings with the Pacific Commander — a standing, senior-level forcing function to review progress on adoption, break through obstacles, and align resources against the theater's requirements at a tempo that annual budget and requirements processes cannot match.
The test-bed proposal was paired with a demand on the military services: they must change their operating concepts to align with the Pacific Command's concept of victory. Rather than each service optimizing its own concepts and programs, and leaving the Combatant Commander to assemble them, the services' concepts — and by extension their force design, training, and investment choices — should be derived from, and be accountable to, how USPACOM and the Joint Force intends to fight and win. Participants acknowledged that deriving the services’ force design, training, and investments from a combatant commander’s concept of victory inverts the customary relationship between service programming and theater warfighting. It was noted that this is not a change achievable by memorandum, but requires a significant renegotiation of authorities between the services and combatant commands.
The proposal drew constructive challenges. It was argued that CENTCOM and EUCOM must also be at the table, given the active wars in Ukraine and Iran.
Nuclear Mission
While the roundtable discussion focused primarily on conventional capabilities, some of the participants also underscored the importance of examining and accounting for the vital nuclear mission and associated evolving strategic environment. It was noted that the intersection and the potential interplay of nuclear capabilities, agentic AI, and automation is extremely complex, with many unknowns. And while nuclear operations and decision-making must retain a direct human in the loop at every step, a suggestion was made that there is room for agentic AI to assist the nuclear enterprise — with early-warning fusion, force posture and readiness tracking, nuclear command and control (NC3) system health monitoring, logistics support for the nuclear triad, and simulation or red-teaming of adversary postures.
Conclusion and Next Steps
The roundtable's answer to its central question was clear in direction if demanding in execution. The character of war is changing faster than at any point in living memory; agentic AI will amplify and accelerate that change rather than merely add to it; and the evidence from Ukraine and Israel is that the binding constraints are institutional — law, policy, culture, talent, and process — not technological. Significant reform within existing structures may not be responsive to changes of this magnitude. What may be required is revolutionary institutional change: transformation of concepts, doctrine, organizations, command structures, training, procurement, and leader development, driven prescriptively from the top, tested against the Indo-Pacific theater's concept of victory, and resourced at the tempo of the problem.
The urgency was framed in competitive terms throughout: adversaries are adapting on six-to-twelve-week cycles, the PLA is suspected of already employing agentic systems offensively in the cyber domain, and the window for the United States to establish a clear advantage is closing and windows of this kind, once closed, are reopened only at great cost. The future battlefield will be defined by teams of humans, agentic AI, and autonomous systems of autonomous systems competing against other such teams — and the future belongs neither to humans alone nor to machines alone, but to those who master the partnership between them.
The task before the United States is not to admire that future. It is to build it – first.
Roundtable Participants