Federal Developers Treat AI Like a Junior Engineer

Federal agencies are increasingly adopting artificial intelligence to accelerate software development, but defense and civilian technology leaders say the technology should be treated like an intern that can boost productivity but still requires oversight.

“When I look at AI, I don’t see it very much different than the new intern that we brought in,” said Air Force Sustainment Center Software Directorate CTO Kurt Jarvis at the Carahsoft DevSecOps Conference in Reston, Virginia. “I have not changed the quality that I have in my pipeline gates. I haven’t changed what my requirements are that goes through the pipeline.”

For mission-critical defense operations, integrating frontier models into software delivery does not mean rewriting established quality standards, Jarvis said.

“I’m not changing my [DevSecOps] pipeline. I’m just making the pipeline run more often, and I’m running it faster,” said Jarvis. “If we ensure that we’ve got the right engineering rigor inside of our process, I can take that manual process and accelerate it as many times as we need to.”

Jarvis said his teams increasingly use AI as a programming partner, allowing developers to automate repetitive coding tasks while keeping humans responsible for reviewing and approving the final product.

“We’ve been doing paired programming in development for a while. Instead of paired programming with another person, I’m a pairing programmer programming with an AI agent, “ Jarvis said. “We ensure that we’ve got the right engineering rigor inside of our process and can take that manual process and accelerate it as many times as we need to.”

Other agencies are taking similar steps to ensure AI operates within controlled development environments.

At the Naval Surface Warfare Center Dahlgren Division, developers use an internally hosted Azure OpenAI environment so engineers can leverage AI without exposing sensitive code or weapons data to public models.

“We stood up our own Azure OpenAI internally to our environment that has guardrails around it,” said the division’s Software Modernization Lead Leroy Mrozowski. “So we know it’s not getting out to the open internet.”

Mrozowski said developers remain the final authority over every AI-generated recommendation. He said teams reject about 70% of AI-generated code before they reach production.

Government Accountability Office (GAO) Director of IT and Cybersecurity Kevin Walsh said that developer pushback is essential against synthetic hallucinations.

“I actually like rejection rate as a metric,” Walsh said. “If you have a 100% acceptance rate, that’s going to be some ugly code.”

Officials also stressed that AI cannot make up for weak cybersecurity. Microsoft Principal Solutions Engineer Kevin Whittle said agencies must establish secure development pipelines before introducing AI agents, not after.

“You still have to have that [system] in place before you do any of those,” Whittle said. “You still have to have it in place before you bring [AI agents] in. If you don’t, you’re just going to bring in more problems than you’ve solved.”

Walsh added that agencies face pressure to adopt AI quickly because adversaries are already using the technology. He cautioned, however, that speed cannot come at the expense of oversight.

“I don’t know that oversight is keeping pace with these frontier models as they are coming in, and I think that’s a mixed blessing,” said Walsh. “We have to graph these models in and use them as quickly as we can because the adversary is using it.”