The secretary of defense’s November announcement on acquisition reform laid out an unprecedented vision across the entire acquisition spectrum to improve the delivery speed of critical capabilities to our warfighters. One foundational pillar is the implementation of a commercial first policy, which laid out several initiatives: leveraging existing but less often used acquisition authorities, maximizing non-traditional contract vehicles, and increasing the use of more flexible solicitations. Most critically, it drives a commercial solution-first paradigm. If the immediate commercial solution isn’t sufficient, acquisition officers then find ways to modify it. If that approach does not work, then and only then do they pursue development.

The spirit of commercial first is to bypass the uncertainty of development and get a sure thing out to the force in a relevant timeline, even if it does not meet every minor requirement. However, as Thomas Sowell said, there are no solutions, only tradeoffs, and a commercial-first strategy poses its own challenges — for instance, costly licensing fees and limitations on data rights in software purchasing. Based on the realities at the time of the announcement, these tradeoffs were worth it given the then-understood alternative of long development lead times.

Much has changed and continues to change since the secretary’s announcement. Businesses have started to see exponentially growing capability in the leading large language models, industry adoption of new AI-optimized interaction protocols, and the proliferation of agentic enabling tools that astronomically reduce the barriers to developing the software they need.

There is no question the department should continue the pursuit of commercial software, but it owes a more critical look at what exactly it should be buying commercially and how it can leverage these capabilities to serve its own internal needs. The department should pursue the commercial procurement of infrastructure, agentic tools, and model access for its existing internal software development organizations. If done, it could build nearly all its software systems and applications in-house with unheard-of responsiveness, full government ownership, and at a fraction of the cost it spends today.

My Agentic Journey

While various military services continue their initial rollout of AI model access through simple chat interfaces, I had a rare opportunity to use true agentic AI that can plan and execute work with data and tools outside the chat window. I was one of a few hundred soldiers and civilians who took part in a pilot program run by the Office of the Army Chief Information Officer and the Office of the Assistant Secretary of the Army for Acquisition, Logistics, and Technology. The pilot gave participants access to an accredited development environment, agentic infrastructure, and a wide variety of AI models. We were allowed to leverage these capabilities to improve our organizational workflows. I chose to create a data structure of acquisition-related U.S. laws, policies, and regulations optimized for agentic AI retrieval, a foundational capability that would support numerous follow-on projects.

I built the pipelines to ingest data from the authoritative sources, pulling in relevant U.S. laws, contract regulations, department policies, and Army regulations. The corpus comprised approximately 60,000 equivalent pages of text, broken up into large paragraph- to page-sized chunks to ensure full legal context when retrieved, and structured around a relational graph to establish document hierarchy. I then wrapped this data structure in a model context protocol server, a recently created open-source standard optimized for agentic context retrieval, making it accessible to any compliant agentic host.

The entirety of the project required 9,500 lines of new code. Using the Constructive Cost Model, a standard tool used in government software cost estimation, this prototype would have taken approximately 6,000 to 7,000 engineering hours, or 12 to 13 months with a four-person development team. Using blended loaded rates derived from listed General Services Administration Labor Categories, this initial prototype delivery would have cost the government between $840,000 and $1.2 million if solicited through a comparable Army small business prototype award.

I completed this project over two weekends, consuming less than 20 hours of my focused attention. Besides my time, the project consumed roughly $1,700 in AI model inference token costs.

I was also able to document and iteratively update the entire project scope, process, architecture, inventory of tools, and deployment instructions throughout the duration of the project. These details were captured in files written in plain English that a non-coder could understand. I then handed this project to a teammate for sustained use with his own agentic tools, even though he was never part of the development process.

I describe all of this, not to claim a unique skill, but as a real demonstration of the sheer power at our hands through agentic AI. These capabilities are turning elite coding into an ubiquitous commodity and delivering it at a speed previously unfathomable just a few months ago, for pennies on the dollar. All of this should, for the betterment of the department, fundamentally change the software paradigm of what it buys, what it makes, and how it makes it.

Leveraging Industry for the Agentic Foundation

In the first quarter of 2026 alone, a total net of $178 billion in new venture funding was raised across the top-tier AI model companies. This is over three times the $58.5 billion the Department of Defense requested for the entire Fiscal Year 2027 budget for Artificial Intelligence and Combined Joint All-Domain Command and Control procurement and investments. With this deluge of resourcing, premier AI models have seen exponential capability growth at monthly cadences. The military services do not have the time or resources to build models with matching capability and should instead buy access to commercial models and infrastructure, procuring that access in a smarter manner.

The department’s first priority should be the procurement of cloud-based infrastructure and services — and, where necessary, on-site hardware — to establish software development and production environments. Historically, bespoke program offices have procured these services piecemeal to support narrow program needs. The result was the duplication of managing costs and compliance burdens. Departmental procurement of these foundational services consolidates the costs across the enterprise. It also reduces the per-program overhead and gives the workforce a common, accredited foundation to build on.

The next commercial procurement focus should be agentic harnesses, the software that allows AI models to interact with external tools and resources. The department needs many solutions for many users: from the regular service member who might need a more chat-like and curated interface to the software developer who needs more transparency and customization to build enterprise-level systems. These harnesses should be AI model-agnostic, capable of leveraging the best capabilities of the day. The rapid cadence of AI capability growth argues against relying exclusively on any one model, lest the department bet on a loser. The current benchmark performance trends illustrate this dynamism, further demonstrating the need to remain agnostic.

To stay model-agnostic, the department should avoid direct procurement of model access and instead leverage AI gateway services. These services allow access to various models through one single endpoint, incorporating consolidated monitoring, compliance enforcement, and flat rate pricing. Through these services, users can apply the best model for the best task through a single connection. As model capabilities emerge and grow, the department can update its workflows by selecting the best model at that given moment. There is, of course, additional overhead in using a gateway service, but the premium is a hedge on the volatile growth of capability. Gateway services provide the greatest selection of options, reduce the cost of transitioning between models, and place the technical burden of new model integration on the service provider.

As these capabilities continue to grow, the department cannot ignore strides within the open source community. Open source models such as Google’s Gemma series, agentic harnesses such as Hermes Agent and Open Code, and supporting tools such as Open Notebook and Honcho offer great capability with no licensing costs. Some will argue that open source code poses security risks, but that is true for all software, commercial or government-built alike. The department should treat these tools with the same scrutiny already applied to any software: code scanning, red teaming for vulnerabilities, and pulling from internally curated repositories rather than public libraries. By its nature, open source code allows what commercial software cannot: full, transparent review down to the source code level, an access not available from proprietary products whose lowest-level code stays walled off from users. These reviews cost hours and people — as onboarding any solution does — but the output is a more diverse toolset for the force, free of recurring licensing, and a means to rein in the costs of paid services.

How should the department acquire these commercial services? Simply leverage the remaining initiatives outlined in the commercial first policy. The little-used Acquisition of Services pathway is particularly promising. This authority has the lowest barrier to entry and is built for large enterprise-wide contracts spanning many users and organizations. Commercial Solutions Openings can be used for persistent solicitations, reopening throughout the year, so industry can submit its best current solution. From those submissions, purchasers should onboard capabilities into multi-award contracts, with pre-negotiated pricing and pay-as-needed ordering. Each contract vehicle should carry multiple options: several cloud and storage providers, gateway services, and agentic harnesses. Only by maintaining diverse selections will industry be incentivized to remain competitive on price and hungry to self-improve, or else risk being discarded for a better alternative.

Where Commercial-First No Longer Applies

With the infrastructure and AI capabilities procured, the force would be empowered to create all the applications to meet its exact process needs with no perpetual licensing costs and no proprietary limitations. If the department does choose to buy an application or business system commercially, it should be priced based on a new paradigm: The agentic technology of today has made it cheap and fast for the services to develop their own software.

Any closed-system solution that requires full adoption into its environment to see value deserves no consideration. Software needs to operate in a vast ecosystem of other applications and external databases. A capability without AI agents that can interact with external systems should not warrant consideration. Leaders should remain ever vigilant of silver-tongued executives trying to sell glorified chatbots wrapped in false claims of agentic capability.

More controversially, the services should plan for their eventual departure from large single data platforms. These systems did initially serve a purpose, allowing access to and building associations between authoritative data sets previously housed in bespoke business systems and locked behind layers of bureaucratic processes to connect. The aggregation of the data then supported critical holistic warfighting and business applications within the platform. However, these systems come with significant costs, and their usefulness for data stores and ontological mappings will be unnecessary due to the advancement of new agentic protocols for interaction and data retrieval. In the FY2027 budget request alone, the department has earmarked approximately $2.3 billion for the licensing and services of the Maven Smart System and the Joint Fires Network. That figure does not cover the various enterprise data platforms across the department, all adding significant annual costs. Annual price tag aside, establishing the connections between systems and platforms, mapping the data ontologies, and building the needed visualizations consume significant labor hours. The opportunity cost of transitioning to another platform becomes so unpalatable that the department effectively locks itself into a single solution, surrendering any leverage and buying power, and violating the open architecture principles it has set forth.

These high annual costs and vendor locks can all be avoided due to the innovations of new agentic interaction protocols — specifically the Model Context Protocol — that have been adopted industry-wide, including by some of the large data platform vendors. With these new open standards, the department can build its own Model Context Protocol servers around all its authoritative data sources. This allows data retrieval from any application that has an agentic AI capable of making calls to the servers. Data discovery and retrieval occurs in real time, negating the need for costly system-to-system integrations.

Organizations would be able to leverage the applications and visualizations of choice in this new architecture paradigm, to include using existing systems if they choose, but divested of the costly and duplicative data storage back end. The department would finally realize the Data Mesh vision it has been pursuing for the last several years: Data would live in its functional domain system via the Model Context Protocol servers and be accessible as a product through agent-to-agent interactions. Being a relatively new protocol, there are, of course, areas of concern to be worked through, as described in the National Security Agency’s publication on Model Context Protocol design considerations. Conscious development decisions and deliberate security controls — most notably the department’s full adoption of role-based permissioning, extended to the agents themselves — can overcome those concerns. There are additional safeguard considerations such as constraining server tool access, restricting privileges, logging, and more. Constructing and securing these servers will take deliberate effort, but the fruits are worth the labor. To get there, the department should pilot the implementation of the Model Context Protocol architecture for select enterprise systems, and if proven successful, expand adoption across the remaining enterprise, operational, and tactical systems. With this architecture in place, the department would truly have a plug-and-play open architecture ecosystem, empowering it to claw back choice and buying power and stop paying recurring licensing fees for capabilities it can stand up at a fraction of the cost.

Arming the Developers We Already Have

These powerful tools are only useful if partnered with knowledge and discipline. Although the department wants to enable its workforce to automate their tasks and workflows at the edge, it still needs developers knowledgeable in good software practices for building data structures and software for enterprise use. Too many news stories detail failures where agentic applications lacked the right safeguards and permissions, causing detrimental destruction.

But the department already has organizations focused on software development. The Army alone has the Science and Technology Centers, Army Software and Innovation Center, Army Software Factory, and various Acquisition Program Offices. If these organizations were armed with these tools, they would match industry in their development capability and generate useful software with full government ownership. To maximize value generation, the department simply needs to introduce basic market forces of internal competition. It can do this by using reimbursable models, reducing the barriers between cross-service interactions, and producing a central system for program offices and organizations to submit work requests that these development organizations bid for. This approach incentivizes efficient value creation, fueled by internal cross-service competition, without needing to dictate specific personnel reductions.

I have no illusions that a development team armed with these tools can summon a business system into existence by vibe-coding. What they can do is stand up an agentic software development pipeline that aids and expedites every phase, from breaking down customer requirements to their technical tasks through development, test, cyber compliance, documentation, deployment, operation, and monitoring. Each phase should be gated, inserting human review at critical points. Documentation becomes living because reviews and revisions are built directly into the workflow. Keeping architectures, software bills of materials, and runbooks current becomes almost trivial. Agents are disciplined by grounding them to those artifacts, forcing an agent to check the code against acceptance criteria before moving it to the next phase. This cuts the risk of hallucinations and undisciplined design. But the real payoff is the ease of sustainment. Because the history and design live in plain English markdown files, continued development can transfer easily between teams — even after gaps in time — to people or agents who were never part of the original build.

The department needs to set conditions for these internal development teams to jump right into work and deliver value when called upon. It can do this through the establishment of pre-accredited development and production environments and a streamlined onboarding process for its internal developers to enter and begin work. These environments should already have the availability of approved software tools, including agentic infrastructure and model access — negating months, if not years, of lead time just to gain security accreditations before building and deploying software. The department is conducting other efforts to rethink and expedite this accreditation process, but those initiatives are externally focused on outside development. We need the same urgency turned inward. Pre-accredited environments with streamlined onboarding for the internal workforce would compress software delivery timelines more than any vendor-side reform and let the department’s existing development organizations start delivering value on day one, rather than after a 6-to-18-month authorization wait.

A Military Revolution within Reach

Agentic AI will have sweeping societal, economic, and military impacts, but I truly believe this disruption can be for the betterment of the department. Through these new capabilities, it can more efficiently achieve its charged duty: to man, train, and equip our forces to allow them to fight and win America’s wars. Realizing these benefits requires the reevaluation of its policies, even its most recent publications. It needs focused commercial procurement to give its forces the necessary tools. It should leverage its internal development capabilities and incentivize them to maximize value generation. It needs to break service silos and establish environments for immediate capability development and deployment. Finally, it should lean forward on the adoption of new architectural paradigms. Each of these efforts will pose its own challenges, but the fruits they will bear are far too sweet to ignore.

Jacob A. Green is a U.S. Army armor and acquisition officer currently serving as the executive officer of the Chief, Automation and Data Integration Office within the Office of the Assistant Secretary of the Army for Acquisition, Logistics and Technology. The views expressed are those of the author and do not necessarily reflect the official policy or position of the Department of the Army, the Department of Defense, or the U.S. government.

**Please note, as a matter of house style, War on the Rocks will not use a different name for the U.S. Department of Defense until and unless the name is changed by statute by the U.S. Congress.

Image: Ashley Low via Wikimedia Commons