Behind the Front Line

Agentic AI—which combines the AI reasoning abilities seen in the LLMs that underlie popular chatbots like ChatGPT, Claude, and Gemini with access to physical and digital tools—is a major growth area for some of the largest AI companies. OpenClaw, an open-source scaffold that allows different LLMs to task an AI agent, received widespread attention for fueling numerous desktop use cases, from managing emails and calendars to shopping online.⁴ OpenAI, Google, and Anthropic, among others, have invested in agentic AI tools built around their flagship LLMs.⁵ These tools streamline workflows and have direct parallels to military use cases in maintenance, sustainment, and planning. As the Joint Force looks to integrate agentic AI, some of the greatest near-term wins may be in these and other support areas, rather than lethal autonomous weapons, which often dominate public discourse—not only because they can deliver meaningful benefits sooner but also because they can give the Department of War (DoW) time to better understand how increasingly agentic systems fit within international humanitarian law/law of armed conflict and existing policies and regulations, such as DoDD 3000.09, Autonomy in Weapon Systems.⁶

This report focuses on three support functions that may benefit from the integration of agentic AI: supply, predictive maintenance, and planning. Although the DoW faces unique challenges, each of these areas has clear analogs in industry, from which the department may draw best practices. In some cases, the department has adopted technologies that will be critical enablers for agentic AI integration, such as additional sensors for inventory tracking, machine learning algorithms for predictive maintenance, and real-time data fusion dashboards for situational awareness. In other areas, the authors identified deficiencies in deployment, such as disconnected business systems, outdated policies, and cultural pushback. If the department hopes to match the increasing speed of operations from greater autonomy on the battlefield, greater autonomy in these support functions will be critical as well.

This report is divided into three sections. The first discusses the use of agentic AI in supply, where autonomy can enable improved tracking and routing of materiel. The second section moves to predictive maintenance, where AI tools can enhance the complete pipeline from detecting faults to diagnosing failures to repairing systems. The third section focuses on planning, where AI can support faster, iterative analysis of alternatives and greater adaptability to shifting circumstances during contingencies. The paper concludes with analysis about the potential of agentic AI for the department and recommendations for enhancing adoption.

Supply

Supply is a critical enabler—and potential bottleneck—in a modern operational environment where forces will be dispersed geographically and supply lines will be contested.⁷ As a military function, supply extends from determining materiel needs to procuring, distributing, and maintaining supply stocks.⁸ To improve efficiency and manage the risk of a near-peer adversary threatening supply lines, future operations require a shift from today’s “pull” system, where sustainment is primarily driven by units placing supply orders, to a “push” system, where unit needs are anticipated and supplies are delivered semi-automatically.⁹ This requires a move toward what the Army calls “precision logistics,” where sensors and AI-enabled tools are used to predict materiel needs.¹⁰ To achieve the goal of resilient supply stocks, the U.S. military will need to work in peacetime to implement the tools necessary to shift supply from a reactive, human-driven present to an agentic future, anticipating needs and proactively pushing materiel to warfighters.

The Joint Force’s supply system, even in relative peacetime, is massive. Through its online system, the Defense Logistics Agency (DLA) manages an estimated 10 million digital invoices and orders per week.¹¹ Once these orders are processed, U.S. Transportation Command (TRANSCOM) manages delivery. In 2023 alone, TRANSCOM transported an estimated 331,000 tons of cargo by air and 13.1 million square feet of cargo by sea.¹²

In a future high-intensity conflict, the scale and pace of military operations will strain supply lines, requiring improvements in automation and predictive analytics to ensure forces continue to receive critical supplies like munitions, spare parts, and fuel. In peacetime, DoW can develop and test the AI-enabled tools needed for resilient supply lines. In wartime, supply operations will have to be dynamic. Static plans that are too brittle to adapt to changing conditions in real-time will not succeed in a contested environment. Adversaries will get a say, too, in undermining supply operations. Agentic tools, consisting of human-machine feedback loops—integrated and tested in peacetime training and operations—provide a means to overcome both the volume of datapoints involved and the flexibility required to leverage them.

An Agentic Vision for Supply

Thus far, efforts to establish a “push” supply system have focused on three critical enablers: real-time information on all supplies, a unified supply order system, and prediction. Each of these elements is essential for autonomous agents to be integrated into supply operations, as they provide the data and interfaces necessary to predict warfighter needs and distribute supplies to them. A common operating picture provides logisticians with a live view of the location and condition of available supplies, so replenishment can be optimized. For example, the Navy is beginning to use radio frequency identification (RFID) tags to locate supplies within shipyards and warehouses, enabling high-fidelity tracking of existing supplies and more efficient planning of future storage.¹³ A unified system for orders allows efficient submission and tracking of supply requests, even with DoW’s hundreds of thousands of suppliers, which enables improved efficiency and the elimination of redundant orders. Since the mid-1990s, the DLA’s Defense Automatic Addressing System (DAAS) has connected military customers with suppliers from both government and industry, ensuring unified data formats and supporting an average of 10 million requests per week.¹⁴ Finally, demand forecasts based on historical data, such as that stored in DAAS, allow enhanced planning for potential needs and reduce the time troops are waiting for resupply by pre-ordering things that they will likely need. For theater-scale planning, tools like DEFCON AI’s Artiv support planners by modeling and simulating potential courses of action for resupply operations.¹⁵ Demonstrated as a part of the Talisman Sabre series of exercises, Artiv uses mathematical optimization to recommend plans that satisfy a set of user-specified constraints within minutes.¹⁶

While service-wide and department-level systems for supply common operating pictures and order fulfillment exist, they remain largely siloed and limited by human-driven processes, such as manual data entry. Efforts to develop predictive and dynamic sustainment tools remain experimental at this point, with several pilot programs underway across the services.

With the critical role that contested logistics will play in a potential U.S.-China conflict, AI agents can not only improve the efficiency of individual supply efforts but also support supply as a first-order consideration in planning itself. Downstream effects, such as predicted supply needs beyond the usual planning horizon, can be factored into the planning process automatically by agents. Further, with the support of modeling and simulation tools, existing operational plans can be automatically evaluated for feasibility and iterated on well before a conflict develops. Intelligence can support automated red teaming of plans against adversary capabilities. Having agents develop and cull supply plans automatically to find potential solutions could support efforts to pre-emplace materiel in theater by recognizing potential bottlenecks.

Using critical data enablers to provide visibility into existing stocks and historically informed predictions about future supply needs, AI agents can quickly develop supply plans that consider both pressing and future needs. While logisticians already have access to the data these agents would leverage, they lack both the time and collective processing power necessary to produce and dynamically address supply requirements. In peacetime, a shift to agent-driven processes will drive greater efficiency. In wartime, it will enable the military to keep its forces supplied and adapt to changes in operational circumstances imposed by a contested environment and changing military demands.

Maintenance

Maintenance—the range of actions from inspection to testing to repair that keep materiel operable—has become a weak point in ensuring readiness for DoW as platforms have aged and grown more complex to repair and costly to operate.¹⁷ DoW spends an estimated $90 billion per year on planned and unplanned maintenance for ground, sea, and air systems, yet readiness levels across the Joint Force remain persistently low.¹⁸ While routine maintenance over the course of a military platform’s lifespan is expected, unplanned maintenance driven by parts shortages, insufficient repair facilities, and personnel shortfalls causes unexpected financial costs and hurts readiness by leaving platforms unfit for deployment.¹⁹ Moreover, when necessary maintenance is delayed due to inefficiencies in the current system, platforms are left to deteriorate further while awaiting repairs, creating backlogs for limited maintenance facilities, spreading delays to other platforms.²⁰ For instance, a consistent lack of trained personnel to support specialized systems, such as the B-2 Spirit, has made qualified maintainers a precious resource that must be optimized to ensure support for mission needs.²¹

Maintenance can generally take one of two forms: proactive or reactive.²² Proactive maintenance focuses on addressing potential causes of failure before they impact functionality, while reactive maintenance responds to failures after they occur.²³ Often, proactive measures support greater availability of systems and reduce costs, as allowing parts to fail can create higher costs in money and man-hours and have cascading effects on other subsystems.²⁴ Recognizing the value in improving efficiency, DoW policy has mandated that the services implement more proactive measures—dubbing them “Condition-Based Maintenance Plus” (CBM+)—since the early 2000s, though to little demonstrable effect thus far.²⁵ CBM+ provides a key stepping stone to agent-driven maintenance, where AI agents will use sensor data and predictive analytics to take actions in the maintenance workflow.

While the mandate has existed for more than two decades, CBM+ remains critically under-implemented across the department. Properly realizing a CBM+ approach requires a complex series of steps and interconnected resources to truly enable cost-effective maintenance that improves readiness levels.²⁶ For example, sensors must either be installed on systems during manufacturing or retroactively installed to provide data.²⁷ After a failure is predicted, root cause analysis is needed to determine where the failure originates to avoid damage occurring again in the future, along with a remediation plan informed by system specifications and manuals.²⁸ Work and operational schedules must be compared to determine the availability of qualified maintainers and deconflict other repair activities, potentially deferring some maintenance based on mission needs and the predicted impact of any failure or combination of failures, such as compounding failures that may originate with one part but place additional stress on another.²⁹

The analytical burden of managing even a single platform’s maintenance pipeline is significant. If limited to human-driven efforts, the DoW would be hamstrung in implementing these processes across the Joint Force. To alleviate the burden on human maintainers, the Pentagon has increasingly embraced automation to improve efficiency and outcomes. In the future, this automation will need to be extended to manage potential knock-on effects from the complex interplay between systems, where proactive or delayed repairs may have unintended consequences on operational planning or maintaining other systems.

Maintenance today remains mostly reactive, with additional scheduled maintenance performed based on static timetables that do not factor in the current state of a particular platform. Attempts by the services to broadly integrate predictive maintenance practices have faced pushback.³⁰ For one, maintainers and leaders often have an entrenched “fly to fail” mindset, performing repairs only after components have already broken, or performing preventive scheduled maintenance uninformed by the system’s actual condition.³¹ Policies for organizations like the DLA have enabled this, requiring, for example, that broken parts are delivered before new or refurbished ones will be issued.³² Pilot predictive maintenance projects have required special dispensations to allow parts to be replaced when they are projected to fail, rather than having already done so.³³ Whether the U.S. military moves to embrace automation or genuinely implements agentic approaches to maintenance, the culture of human maintainers and operators remains a critical barrier to realizing the readiness improvements these tools can provide.

Transitioning to fully agentic maintenance practices will also carry a high up-front cost, requiring system health monitoring equipment to be installed on new and legacy systems; retraining operators and maintainers; and integrating software and infrastructure to transmit, receive, and process data.³⁴ With replacement parts, as well as qualified maintainers’ time, in short supply, it may not be possible to complete all needed repair tasks without broader changes to supply chains and the military maintenance workforce.³⁵ Much like the cultural barrier previously discussed, agentic tools cannot resolve persistent spare parts shortages and maintainer workforce gaps on their own.

The availability of the data necessary to execute an enhanced predictive maintenance regime also presents a significant challenge, as multiple participants highlighted during expert workshops hosted by the Center for a New American Security’s (CNAS) Defense team.³⁶ Maintenance data, like previous repairs, specifications and manuals, and repair schedules, may be spread across multiple systems, necessitating human operators to navigate between sources in developing a maintenance solution.³⁷ Further, the Government Accountability Office has reported on the lack of access to sufficient intellectual property from manufacturers to enable repairs.³⁸ This information is critical in determining how to perform a repair. Analyzing diagnostic data and predicting system failures are only the first steps in the maintenance process. Additional technical data is required to close the loop from discovering a system failure to identifying a fix to supplying replacement parts and completing repairs. Agentic tools will help human maintainers make sense of and act on the vast amount of data that exists. However, without the ability to access and interrogate all the necessary data, they will not be any more effective than the current human-driven approach to maintenance.

Maintenance strategies must take a holistic approach to tackle these roadblocks, considering operational priorities, optimizing limited resources to align with military objectives, and weighing the upfront cost of integrating new techniques with downstream savings. This “art”—assigning scarce resources and aligning them to mission needs—is currently a human-driven process. But continued developments in AI-enabled logistics tools can increase the pace and adaptability of maintenance planning.

An Agentic Vision for Maintenance

Agentic tools that leverage real-time operational, supply, and maintenance data offer the department the means to automate the most time-intensive aspects of the maintenance cycle. Human-driven efforts to diagnose maintenance needs, requisition parts, and perform repairs create a bottleneck of inefficiencies that moves too slowly and cannot keep pace with the Joint Force’s operational needs. The department has pursued data-driven processes to reduce unplanned and unnecessary maintenance and improve readiness across the Joint Force in ways that begin to leverage the data and computing resources increasingly available to it.³⁹ These processes use large amounts of data from individual platforms, along with AI, to identify when, what, and how maintenance is required. The agentic future of maintenance extends this further, allowing AI agents to automatically complete some of the tasks currently managed by maintenance personnel, such as diagnosing part failures, ordering replacements, and scheduling repairs.

AI agents will take responsibility for monitoring the health of weapon systems and platforms, detecting failures, diagnosing malfunctions, requisitioning the materials needed to mitigate them, and scheduling the repairs. These functions will reduce the degree of human effort and time needed to conduct these actions, greatly speeding weapons and platforms back into service.

As a military function, maintenance is not just the act of repair itself. Implementing responsive and robust maintenance requires sequencing several operations that are spread across disparate systems. While these processes may vary across industries and military services, the authors adopt a simplified model that steps through each of the common stages encountered in maintenance. First, data must be collected from sensors to construct a picture of the system’s health (sensing). This data feeds into system models and/or analyses based on historical fleet data to determine whether there are any existing or projected system failures (detection). Once a failure is detected, additional information, including system specifications and manuals, must be analyzed to determine what the root cause of the failure is or will be, so it can be remediated (diagnosis). Required components to complete an expected repair must be located, anywhere from the spare parts kits deployed with a unit up to national-level warehouses or suppliers (requisition). Finally, operational and maintenance schedules must be deconflicted to ensure that a maintainer with the right skills for the repair is available at the right time, considering other mission needs (scheduling). With greater integration of logistics data sources and decision-making aids, each of these steps can use AI agents to support warfighters, maintainers, and leaders.

Achieving the vision of an end-to-end agentic maintenance workflow will require augmenting each of these stages with agentic capabilities. This will enable critical decision-making at each stage to be automated, with human intervention to avoid potential failures. Agents must be deployed at the platform, fleet, and command levels, coordinating with one another to manage and deconflict repairs. While each of these stages are nominally discrete, agents will need to intercommunicate to account for the knock-on effects of their decision-making. For example, a requisition agent’s timeline for ordering a spare part may directly impact a scheduling agent’s ability to plan a repair and find available staff. The interconnection of these agents will require careful orchestration and rely on engineering techniques for multi-agent systems. Each of these stages is considered in turn, as well as the overarching agent orchestration problem.

Figure 1: Phases of the Maintenance Process

Sensing Agent

Generating and collecting system health data is a critical component of maintenance, one where intelligent automation is valuable.⁴⁰ Sensors may generate thousands of datapoints per second per platform, which can be collected and analyzed to generate a picture of a system’s health.⁴¹ Agentic tools enable human maintainers to make sense of an overwhelming volume of real-time data, scrutinize the quality of that data, and prioritize actions for maintainers to take based on operational needs and the availability of parts.

Data quantity does not outweigh quality, though any method that produces more accurate and timely assessments could be valuable.⁴² These assessments, whether by human maintainers or agents analyzing sensor data, can find undetected maintenance issues, which can extend backlogs.⁴³

Storing, cleaning, and managing this data is a significant feat unto itself. AI agents equipped to perform automatic quality checks on incoming data can prevent poor data from triggering or obfuscating downstream maintenance recommendations. They could also alert maintainers to potential sensor issues, distinguishing faults in sensors from other hardware issues. As the department has recently recognized, secure, interconnected, and distributed data stores are required to ensure that stakeholders can access data in a timely fashion.⁴⁴ With the rise of data poisoning by adversaries as a threat to AI models, databases also become a potential attack surface.⁴⁵ AI agents and operators must provide quality control to ensure that data can be used for evaluations further down the pipeline. As AI systems are “garbage in, garbage out,” data errors can lead to false recommendations from AI that may mislead operators if they are over trusted.

Diagnostic Agent

As part of an agentic approach to maintenance, an AI agent would first combine sensor data to determine whether a failure has occurred or is imminent. Historical data from other systems could be used to identify patterns that preceded previous maintenance issues, such as unusual vibrations or operating temperatures within machinery.⁴⁶ Physics-based models could even predict failures by combining data about the system’s current state with simulations of future use.⁴⁷ With the overwhelming amount of available data, agents extend human capabilities by performing analysis at machine speeds, correlating data in ways that humans may be unable to. Once faults are detected, an agent can alert maintainers to the anomalies and diagnose their root cause.

Root cause analysis requires AI agents to develop an understanding of the system and its interconnected components—something that they could do much faster than humans who require extensive training.⁴⁸ While performance may drop in one subsystem, the failure may be caused by other upstream components. For example, a change in a vehicle’s fuel efficiency may not indicate fuel or engine issues but rather that a tire is flat. This requires agents to review technical specifications and manuals, which may consist of a mix of structured and unstructured data. Researchers have demonstrated in recent years that LLMs can excel at analyzing this sort of mixed data, which suggests that diagnosis is one step in the predictive maintenance pipeline where agentic tools with strong reasoning capabilities will be beneficial.⁴⁹

Once the cause is identified, mitigation or repair options can be determined and presented to operators for approval. This could include temporarily removing the platform from service, if necessary, or requisitioning spare parts to have on standby for future scheduled maintenance.

Requisition Agent

If a repair need is identified, the next step is to locate and request any necessary spare parts. If these are co-located with the platform, such as in spare parts kits distributed with deployed units, they may be easily utilized for repairs. This is often not the case, however, especially for high-end platforms with critical parts that are expensive, sparse, and difficult to store with units.⁵⁰ An AI requisition agent must be able to identify the location of equipment and, if necessary, escalate its search to other units or commands.⁵¹

Distributing equipment, especially sparse and expensive equipment, is a complex process that requires balancing competing priorities.⁵² Logisticians, in concert with agents, will need to regularly weigh distribution of parts at the unit, theater, and national levels.⁵³ An AI requisition agent can weigh these competing priorities, informed by strategic and operational goals and logistical constraints. Suppressed supply lines in a conflict could scuttle agent-generated plans, requiring the agent to run persistently, maintaining oversight from when the request for equipment is generated until it is satisfied. If a part will not arrive quickly enough to complete maintenance before a planned operation, the agent might develop a backup plan or alert leadership.

Scheduling Agent

Finally, a scheduling agent must identify appropriate personnel to perform maintenance, accounting for their availability and competing priorities. Limited availability of qualified personnel can make planning maintenance a challenge, one that has been identified repeatedly in meeting readiness goals for both ships and aircraft.⁵⁴ Distributed operating concepts must also take into account the ability to either move platforms to depots for complex maintenance or deploy maintainers forward.⁵⁵ In a high-intensity conflict, where maintenance may be more frequent due to a higher optempo, this agent must automatically prioritize maintenance to match pressing operational needs.

Maintenance scheduling is itself not a novel task. However, in the context of a high-intensity conflict, the environment, scheduling demands, and parts availability become exceedingly dynamic. For example, diverted shipments may delay the availability of critical parts for a repair beyond the planned maintenance window. Maintenance delays can cause further deterioration, spiraling to create even deeper maintenance backlogs.⁵⁶ Predicting these knock-on effects requires predictive maintenance tools like those in the diagnostic step above. While a maintenance need may not be deemed critical yet, repeatedly delaying it could diminish the readiness of a platform.⁵⁷ The scheduling agent thus requires access not only to information about maintainer schedules and parts availability but also operational plans. Human maintainers will be in high demand in a conflict environment and agentic tools will free them up from the task of scheduling repairs. With this approach, maintainers will be able to devote their energy to performing the maintenance tasks that require human execution while the agentic workflow described above works in concert to enable them to act efficiently and promptly.

Agent Orchestration

Each of these steps and their associated agents must be managed by some overarching supervisory agent or policy, potentially captured implicitly in the design of the agents themselves. Knock-on effects between agents can inadvertently cause problems as they each prioritize their own objectives. For example, a requisition agent needs some insight into scheduling to understand when parts must arrive to perform repairs. The complex interaction of logistical constraints and operational needs must inform each element of the maintenance process.

Competing goals and decision-making logic across agents can introduce conflict, making interoperability a major focus of developing maintenance agents. While each of the maintenance stages outlined here is considered in isolation, for a single repair job, in practice, agents will have to manage many parallel and asynchronous maintenance needs. The agents must account for the interdependencies and conflicts between these needs, responding dynamically as elements of the supply chain shift. For example, delayed delivery of a spare part for one platform—and a concomitant delay in its scheduled maintenance—may open space in a depot and maintainer’s schedule to service another platform. This requires intercommunication between the requisition and scheduling agents to identify the opportunity and ensure maximum efficiency.

Beyond behavioral interactions between agents, access to shared, consistent data is critical to enabling inter-agent cooperation. Stale data can lead agents to establish seemingly correct plans that fail in practice. Managing this risk not only requires the department to ensure that it has the data infrastructure itself but also that operators are trained to identify these mistakes when databases fail to maintain consistency, interrogating plans that may seem correct at first glance.

Limits of Agent-Enabled Maintenance

Agentic tools for maintenance can help to reduce risk, increase readiness, and predict needs, but they are not a panacea for the shortcomings inherent in the Joint Force’s maintenance enterprise. Personnel shortages, supply chain delays, and unpredictable conflicts will inevitably reduce the effectiveness of these tools. Further, predictive maintenance solutions cannot make up for accumulated repair debt and a lack of parts and maintenance personnel. Maintenance backlogs have been noted repeatedly across services and domains, leading some platforms to be prematurely retired due to maintenance cost overruns.⁵⁸ In at least some cases, these cost overruns have been caused by delayed maintenance, where a platform has continued to deteriorate, sometimes due to leadership prioritizing keeping platforms deployed to cope with near-term operational demands and limited shipyard capacity.⁵⁹ Building a robust maintenance system that can support distributed operations and contested logistics requires improvements outside the act of repair itself.⁶⁰

Proactive maintenance, managed by AI agents, can improve the availability of critical systems and better inform decision-making surrounding maintenance activities, but some reactive maintenance will always be inevitable. AI tools cannot predict all possible failures, especially those due to novel adversary actions that are not represented in the historical data used to train machine learning models for failure prediction or simulation. To manage reactive maintenance, AI agents must be adaptable and supply chains must be efficient enough to respond to emergent needs. Planned deliveries and maintenance schedules based on forecasts may be changing constantly, reprioritizing certain repairs based on mission needs. Agentic tools can act as a force multiplier in a dynamic conflict environment where logistics resources and activities are contested. Human maintainers will simply not be able to keep pace with the demands of keeping platforms mission capable.

Operational Planning

In just the first 100 hours of the war in Iran, the United States and Israel claimed to strike more than 4,000 targets.⁶¹ The Joint Force has expended thousands of munitions prosecuting tens of thousands of targets.⁶² For each target, military planners, logisticians, operators, and commanders have undoubtedly reviewed operational plans (OPLANs), air tasking orders, target prioritization and no-strike lists, and intelligence to craft and execute their strategy. Materiel had to be moved in and around the theater, runway and maintenance schedules had to be deconflicted, and strike packages had to be assembled. As the war demonstrates, the pace of conflict between even asymmetric adversaries is increasing rapidly, fueled by artificial intelligence and other advanced technologies.⁶³ In preparing for a near-peer adversary, planning will remain critical to coordinating forces and prosecuting strategic objectives, both before and during a conflict. Agentic planning tools can provide overburdened planners with the ability to maintain military plans during peacetime and enable them to dynamically adjust them to changing operational circumstances during wartime.

Defense planning occurs from the strategic down to the tactical level. The Office of the Secretary of Defense develops the National Defense Strategy (NDS) to both reflect and shape force structure across the services.⁶⁴ The Joint Staff further refines the priorities outlined by the NDS in the National Military Strategy, outlining how the Joint Force will be reshaped to meet strategic objectives.⁶⁵ This high-level planning can have major impacts on force structure but extends further, shaping combatant command campaign plans and OPLANs and eventually affecting individual mission planning. These plans become highly concrete, outlining major logistical efforts to posture forces in theater and prepare them to prosecute operations.

Undertaking the complete pipeline of planning requires a mix of military science and art.⁶⁶ Commanders engage with their staffs to develop a comprehensive picture of the situation in their AOR, review available ways and means, assess risks, and implement strategies to achieve their objectives.⁶⁷ These plans are not static and require regular revisions as circumstances shift, operational objectives evolve, and military capabilities progress. This means that even after the human labor of developing new military plans is complete, an existing plan continues to require planners to commit time to keeping these plans up to date—often through routine but labor-intensive practices such as updating logistics tables, modifying intelligence reports, and adjusting deployment timetables. In-depth planning documentation, like Time-Phased Force Deployment Data, is manually developed and updated, articulating in painstaking detail how troops and materiel will move during an operation.⁶⁸ As the Joint Force moves into the age of artificial intelligence, agentic tools will provide planners with the ability to offload routine tasks and, in cooperation with human planners, ensure that defense plans remain up to date.

The Joint Planning Process moves through a series of steps that allow the commander to understand the operational environment, define the problem the plan is meant to solve, develop a concept and compare it to alternatives, and assess operational risk.⁶⁹ Given the range of scenarios planning is meant to address, planners are trained extensively on the planning process “bible,” Joint Publication 5-0.⁷⁰ Planning tasks are rigorously spelled out as a sequence of steps, but the act of planning itself is often described as an art, requiring commanders to think creatively about the problem they face, assumptions in their thinking, and ways to employ the military means available to them. Headquarters staffs dedicate themselves to fact-checking the underpinnings of plans and revising them based on intelligence on adversary capabilities.⁷¹ The best application of agentic planning tools is not to replace the human-driven “art” of planning but to provide planners and military leaders with faster, more efficient means of developing and maintaining plans. Improved efficiency will lead to more frequent updates to existing plans during peacetime, allowing human planners to rapidly iterate and modernize detailed operational plans as circumstances change over time.

The quality of these plans is inherently linked to the quality of the situational awareness the commander and their staff can create, requiring accurate intelligence about both blue and red capabilities, politics, and incentives. For this reason, especially as an operational environment continues to shift, previously drafted plans can become “stale,” requiring refreshes based on new information.⁷² This can especially be true for the cyber domain, where novel capabilities and zero-day vulnerabilities may emerge unexpectedly.⁷³

During execution, these plans are stressed further. Facing contested logistics with a dispersed Joint Force, carefully laid plans may require rapid adaptation as supply lines become stressed and materiel is expended.⁷⁴ Planners, logisticians, and commanders must work under time pressure to adapt these plans as a campaign unfolds. This additional pressure creates a drive toward “satisficing,” as one former senior commander stated, pushing commanders to pick the first “good enough” plan.⁷⁵ As AI systems continue to advance, developing the ability to reason over longer time horizons and integrate with modeling and simulation tools, military staffs will be able to leverage agentic planning tools to iterate on compressed timelines to create better plans at speed.

The department has made some inroads toward integrating advanced AI tools to support planning at different levels. The Maven Smart System fuses real-time sensor dataflows to provide situational awareness and supports targeting workflows, enabling dynamic mission planning.⁷⁶ LLMs on both unclassified and classified networks can read strategic guidance and doctrine, allowing users to query them as they translate guidance into concrete operations.⁷⁷ These models can also ingest intelligence assessments, providing brief, cited summaries.⁷⁸ As the Pentagon looks to integrate agentic AI into planning, these tools will move from simple summarization to producing concrete actions that planners and commanders can review and implement.

An Agentic Vision for Planning

As agentic AI continues to develop for military planning applications, planners can rapidly explore the space of alternatives for a given course of action, moving beyond the “satisficing” mindset and creating solutions that are not just sufficient, but optimized. As multiple participants in the CNAS Defense team’s workshop series remarked, planning staffs face deadlines, especially during contingency operations, that limit their ability to explore the range of options available to them for a particular objective.⁷⁹ Much as LLMs are already enabling individuals to rapidly iterate on draft documents or presentations, the next generation of AI equipped with external tools may support semi-automated analysis of alternatives through linkage with simulation tools and intelligence sources. An individual planner could iterate on an operational concept with an AI agent, refining her ideas and testing them virtually all in a single workflow.

For concept development and assessment, AI agents can evaluate ideas, working with individual planners to revise and test concepts much in the way entire planning staffs work now. Joint Staff doctrine already notes that, although these steps are nominally sequential, they often occur concurrently to save time.⁸⁰ By directly integrating AI tools that can perform some high-level assessments, planners can save time by offloading much of the modeling and simulation work conducted by human planners to agentic tools. Proposed concepts can be iteratively refined with the AI models, producing strong initial solutions for human planners to adjudicate in days rather than the months-long time frame currently associated with developing or updating plans.

AI agents can speed adaptation of plans as circumstances on the ground change. For human planners, responding to unfolding crises imposes the greatest time constraints.⁸¹ AI agents can not only support crisis response through speed-ups in iterative plan development but could also directly access live data to inform planning efforts based on the most up-to-date picture of the situation on the ground. These tools can support good operational design, using both military and open-source data to perform analysis that has been historically constrained to human intelligence officers, such as assessing the political environment and public sentiment in the area of responsibility.⁸² Further, routine plan updates, conducted in peacetime, can also be automated with human oversight, allowing up-to-date plans to be ready before crises unfold.

Rather than AI agents serving primarily to replace experienced staff, they can be integrated in concert with existing modeling and simulation (M&S) tools and human planners. The interconnection of these three parts—a human planner, one or more AI agents, and M&S tools—can be described as the “triangle of agentic planning” (Figure 2). This approach takes advantage of the strengths of each of the integrated parts. M&S tools can use physics-based models and historical data to test the expected impact of a proposed course of action. AI tools can efficiently orchestrate calls to these M&S tools, which can be computationally expensive to run, produce partial solutions to seed the M&S tools based on historical operational training data, and track the phases of planning through to eventual approval and execution. Finally, the human planner can check for correctness, ensuring plans optimize for the correct variables and rely on valid assumptions, and produce novel ideas that may be outside the capabilities of either the M&S tools or the AI agent.⁸³

Figure 2: Triangle of Agentic Planning

The key differences between AI agents and M&S tools in the planning workflow are the speed and adaptability of agents. M&S tools may produce optimal or semi-optimal solutions but require significant computing time.⁸⁴ For example, routing algorithms for mission planning are guaranteed to produce optimal solutions under a given set of constraints, given enough time. The amount of time these algorithms take to run is proportional to the size of the map they must consider, so larger search spaces can take exponentially longer run times.⁸⁵ AI agents are not guaranteed to produce optimal solutions but may be able to produce several approximately correct solutions quickly, which can feed into and narrow the M&S tools’ search, improving speed. M&S tools are further limited by the variables chosen during their engineering, such as which enemy threats and environmental constraints programmers chose to model. As operational environments change and new threats emerge, re-engineering these tools may be costly, while an AI agent could manage these new constraints more rapidly.⁸⁶ By integrating these tools with each other, planning teams can take advantage of the strengths of both, using M&S tools when appropriate to produce predictable and optimized solutions while generating candidate solutions more rapidly with AI.

Limits of Agent-Enabled Planning

Although AI tools for planning offer potential speed-ups, they still retain many of the same risks publicized by AI researchers in recent years. Hallucinations—when AI systems present false information as fact—will be a challenge to integrating reasoning models in planning tasks.⁸⁷ Relatedly, bias in these systems absorbed from their training data can introduce mistakes or faulty reasoning into plans, much in the way a human planner’s unsupported assumptions can.⁸⁸ Planners must be trained to be aware of these failure modes so they can correct them, understanding how to independently analyze AI outputs based on their own knowledge and experience.

Future planners will need to retain the skills and background for manual (non-AI-assisted) planning to properly adjudicate the options provided by agentic planning tools.⁸⁹ This catch-22, that planners must simultaneously use AI tools that offload some of the cognitive load of planning while retaining the skills needed to perform planning on their own, will require thoughtful re-design of training and military professional education. The Pentagon has considered the potential for skill atrophy with respect to lethal autonomous weapon systems, where operators must be prepared to take manual control of a system if its autonomy fails, but less attention has been focused on this problem for support functions.⁹⁰

A component of training and system design will also have to focus on developing calibrated trust between users and AI agents.⁹¹ Some users may be inclined to trust AI outputs more than they should, overestimating the agent’s capabilities (overtrust).⁹² Others may automatically distrust AI outputs (undertrust).⁹³ Calibrated trust is the sweet spot where users have sufficient insight about a system’s capabilities and risks to know when and how to use AI appropriately.⁹⁴ From an engineering perspective, tools for explainability, like auditable chains of reasoning and calibrated confidence scores for outputs, coupled with logs of tool calls, can provide additional insight to users about how a system’s recommendation or assessment was generated.⁹⁵

Conclusion and Recommendations

Artificial intelligence has the potential to change many facets of future military competition. While popular discourse has focused primarily on the risks and potential of lethal autonomous weapon systems, AI integration into critical military support functions may pay nearer-term dividends for the Joint Force. By integrating agentic AI into supply, maintenance, and planning, the department can reap efficiency benefits that are necessary in facing an advanced adversary.

This is not to say that agentic AI does not carry risks. Hallucinations from LLMs can misdirect military planners, especially if they are not prepared to critically evaluate AI outputs for accuracy. AI systems themselves also become a new attack surface for adversaries, through tactics like data poisoning or adversarial examples. To reap the benefits of AI, the department will need to continue pursuing frameworks and tools for assessing risk, training operators and leadership on the dangers of AI, and developing calibrated trust between warfighters and AI tools.

Recommendations to Enable Agentic AI Implementation

1. Provide training for operators in the use of agentic tools, as well as evidence demonstrating their value in improving operations. Current operators may not have significant exposure to AI and are trained on legacy processes and mindsets that may be directly at odds with these tools, such as the “fly to fail” mindset that has dominated the maintenance community. Training can support greater adoption of agentic tools, especially by demonstrating potential efficiency gains. For training to result in operators embracing agentic tools and best practices, servicemembers must be confident that they are an improvement upon prior human-centric approaches. Training processes, therefore, should emphasize the benefits of agentic tools while also educating operators on their use. This two-pronged approach will help overcome the adoption challenges seen most notably in the Pentagon’s deployment of predictive maintenance practices over the last two decades.
2. Enhance real-time, accurate data collection. To enable operations, especially in dynamic environments, agents will need a real-time view of available supplies, force posture, and orders. Ongoing efforts to collect this information through low-cost sensors like RFID tags should be expanded within and across services. One of the primary roadblocks facing maintenance and supply is the availability, accuracy, and timeliness of the information collected across service-wide enterprise systems. Agentic tools can help overcome this challenge, but they will only be as good as the quality of the data they are provided.
3. Unify data from disparate systems. AI agents contributing to supply, maintenance, and planning will require real-time access to data that is currently housed across disparate systems. To orchestrate efficient operations, agentic tools will need to see not only data in their nominal domain, such as maintenance schedules, but also that of other agents, such as parts requisition orders or operational schedules. These datalinks will have to be bidirectional, allowing agents to identify conflicts and recommend or execute remedial actions, such as prioritizing certain orders or repairs.
4. Establish high-level support for agentic tool adoption with the authority to clear cross-agency roadblocks to integration. With the complexity of supply, maintenance, and planning, elements of several commands, agencies, and services will be involved in any modernization effort. For example, DLA policy stymied initial predictive maintenance efforts by preventing the premature replacement of failing parts. Having buy-in from senior leadership, as well as the authority to change policy, can support implementation through unidentified hurdles remaining from existing processes. By necessity, agentic tools will conduct activities across agencies and organizational charts, working seamlessly with one another to leverage data. High-level advocacy for these tools and the stovepipes they will work to remove will help overcome the parochial challenges that have limited prior efforts to unify data across the Department of War.
5. Identify commercial dual-use agentic tools to provide near-term capabilities. Industry is already deploying agentic tools to support dual-use applications, such as maintenance diagnostics, planning, and execution. Through organizations like the Defense Innovation Unit, DoW should identify these tools and adapt them to Joint Force use cases, rather than relying on traditional development and acquisition pathways for defense-specific software.

Recommendations to Develop Guardrails for Agentic AI Use

1. Phase deployment of agentic tools and maintain ongoing dialogue between developers and users for tool refinement. Agentic tools are unlikely to work without friction during their initial deployment, both as bugs and vulnerabilities are uncovered and users adapt to new workflows and organically develop new uses for agentic tools. Deployment should be phased, with early exposure that allows operators to provide feedback and developers to rapidly iterate. Once initial deployment transitions to widespread enterprise adoption, DoW should implement processes for regularly incorporating operator feedback and insights on incremental improvements.
2. Provide infrastructure for continuous monitoring and maintain long-term support for developers. Agentic tools can enhance efficiency but also introduce a novel attack surface for adversaries seeking to undermine planning and operations. As operational environments change and adversaries adapt, likely using their own AI tools, developers will need to continually refine agentic tools to maintain their resiliency, correctness, and efficiency. This is particularly true for machine learning–powered tools, which rely on recent, representative data to produce correct outputs.
3. Provide top-down support for developing and disseminating agentic tools and encourage ad-hoc experimentation by operators. Operators across distinct areas of responsibility will have shared needs and procedures, such as maintenance planning, which enables agentic tools to be shared across the Joint Force. Simultaneously, the eccentricities of certain operating environments and mission-sets may present specific needs. Operators should be provided, within appropriate constraints, with the flexibility to develop or modify their own agents to fit their needs in an ad-hoc fashion, disseminating their lessons learned back up the chain of command. Bottom-up solutions development will be essential to ensuring that agentic tools remain relevant to operators’ real-world needs. No tool will be perfect when it is initially deployed, and regular refinements will be needed.
4. Perform rigorous testing and evaluation and red teaming throughout each system’s life cycle to understand potential vulnerabilities and limitations in deploying agentic tools. As with human operators in a congested information environment, agentic tools will be subject to exploitation by adversaries who may take advantage of these tools to undermine operations. Tool vulnerabilities should be remediated when possible and, when not, operators should be trained to understand limitations and identify failures as they arise. These lessons learned should be collected and disseminated across users. Adversary efforts to undermine agentic tools will be widespread and constantly evolving. The solution is not to strive for tools that are immune to attacks. Rather, operators should be trained to understand and mitigate those vulnerabilities, and the tools themselves should be regularly refined to account for adversary efforts as new vulnerabilities are discovered.

About the Authors

Josh Wallin is a fellow with the Defense Program at the Center for a New American Security (CNAS). His research focuses on military autonomous systems, including rigorous test and evaluation, responsible AI employment, and collaboration with allies and partners.

Prior to joining CNAS, he was a technology and national security fellow on the Joint Staff at the U.S. Department of Defense. In this role, he assessed the impact of emerging technologies on the future Joint Force. Previously, he worked as a software engineer at John Wiley & Sons.

He holds an MS in computer science from Northeastern University and a BS in computer engineering and Spanish from Iowa State University.

Carlton Haelig is a fellow with the Defense Program at the Center for a New American Security (CNAS). His research and expertise include national security strategy, force design and employment, and military innovation.

Prior to CNAS, Haelig was a postdoctoral fellow with the America in the World Consortium at the Clements Center for National Security at the University of Texas and the Kissinger Center for Global Affairs at Johns Hopkins University’s School of Advanced International Studies. He has held prior research and analysis positions at the University of Pennsylvania Center for Ethics and the Rule of Law, RAND, and the Historical Office of the Office of the Secretary of Defense.

Haelig holds a PhD in security studies from Princeton University where he directed the Strategic Education Initiative at the Princeton Center for International Security Studies. He also completed an MA in public affairs at Princeton, an MA in international security at George Mason University, and a BA in political science and history at Rutgers University.

About the Defense Program

Over the past 19 years, CNAS has defined the future of U.S. defense strategy. Building on this legacy, the CNAS Defense Program team continues to develop high-level concepts and concrete recommendations to ensure U.S. military preeminence into the future and to reverse the erosion of U.S. military advantages vis-à-vis China and, to a lesser extent, Russia. Specific areas of study include concentrating on great power competition, developing a force structure and innovative operational concepts adapted for this more challenging era, and making hard choices to effect necessary change.

Acknowledgments

The authors would like to thank their colleagues in the Defense Program, particularly Dr. Stacie Pettyjohn, Phil Sheers, Molly Campbell, and Delaney Soliday, for their support in researching and writing this report. They would also like to thank all the participants in their AI agents workshop series, who generously shared their time and expertise. They also thank Joe Chapa and Lt Gen Jack Shanahan (Ret.), USAF for their thoughtful review and feedback on this report. This report was made possible with general support to CNAS.

As a research and policy institution committed to the highest standards of organizational, intellectual, and personal integrity, CNAS maintains strict intellectual independence and sole editorial direction and control over its ideas, projects, publications, events, and other research activities. CNAS does not take institutional positions on policy issues, and the content of CNAS publications reflects the views of their authors alone. In keeping with its mission and values, CNAS does not engage in lobbying activity and complies fully with all applicable federal, state, and local laws. CNAS will not engage in any representational activities or advocacy on behalf of any entities or interests and, to the extent that the Center accepts funding from non-U.S. sources, its activities will be limited to bona fide scholastic, academic, and research-related activities, consistent with applicable federal law. The Center publicly acknowledges on its website annually all donors who contribute.