Beyond Automation: How Agentic AI is Reshaping Federal Missions

Federal agencies are testing the limits of agentic AI to boost mission delivery and drive, agency leaders say.

“We’re not looking at the technology itself. We’re really saying, let’s define the problems our agencies are looking to solve, and through that life cycle management of technology, where is [agentic AI] applicable? Where is automation applicable?” said Defense Logistics Agency CIO Adarryl Roberts in May.

In April 2025, Office of Management and Budget Director Russell Vought issued a pair of memos that redefined the federal approach to AI, including agentic AI systems. “Agencies should focus recruitment efforts on individuals that have demonstrated operational experience in designing, deploying and scaling AI systems in high-impact environments,” Vought wrote. The directives emphasized the need to remove bureaucratic barriers and empower Chief AI Officers to act as “change agents,” rather than gatekeepers.

Agencies face hurdles in adopting agentic AI, experts said. Thomas Robinson, COO of Domino Data Lab, told GovCIO Media & Research that trust is paramount for agencies to adopt AI – including agentic applications – into systems.

“Regardless of its if it’s an agent, it’s a chatbot, it’s a machine learning model, it’s a spreadsheet with a linear regression,” Robinson said. “Whatever the application may be, the hard part is in building trust into those systems.”

USPS Eyes Agents for Logistics and CX

U.S. Postal Service CIO Pritha Mehra said that the agency is beginning its agentic AI journey, leveraging agents to personalize experiences for customers and automate software development, during the GovCIO Media & Research’s 2025 Federal IT Efficiency Summit in Tysons Corner, Virginia, in July. Mehra added agentic AI can help USPS become more efficient within supply chains and other complex tasks.

“We’re looking at automating our entire DevSecOps pipeline, but truly, to get to where we want to go, you want to bring in agentic AI that will automate your complex tasks and personalize [customer] experience, which will really help drive the efficiencies,” Mehra said.

USPS is using agentic AI to better understand and leverage data produced by interactions within the service, Mehra said. Agencies like USPS can better serve the mission by deploying AI agents, she added.

“Those organizations that are able to first of all prepare their data, then harness that data with all these intelligent tools,” Mehra said. “These are the ones that are going to win.”

Personal Health Agents and Predictive Care at Health Agencies

The Advanced Research Projects Agency for Health (ARPA-H) has embraced agentic AI as a key focus of its AI strategy. The agency uses a variety of AI models acting as agents to monitor and check one another, according to ARPA-H Director of Data Innovation and Acting CDO Alastair Thomson.

“How do we use AI to keep it tabs on AI?” Thomson said during a recent GovFocus program. “We work with Microsoft, OpenAI and anthropic, because they’re all coming at these things slightly differently.”

The agency is currently using this approach in their chatbot, which functions as an “AI scientist” for program managers and staff. The chatbot, named Grace after Grace Hopper, can detect hallucinated citations, a problem that has even appeared in recent scientific publications, Thomson, said. This is part of a broader effort to use AI to ensure the accuracy and reliability of information, especially in complex scientific fields where a human may find it difficult to keep up with the state of the art, he added.

“We’re using a variety of AI models as agents to check on one another. We’re doing this in Grace at the moment, where we’ve designed an agent to detect hallucinations of citations. There’s been scientific publications recently that shall remain nameless, that had hallucinations in them, which was a bit embarrassing,” said Thomson. “We can’t afford to do that.”

The Department of Veterans Affairs is looking to bring agentic AI systems to augment its health care workforce, Shane McNamee, former health solution architect at the agency, told GovCIO Media & Research at HIMSS earlier this year.

“[With agentic AI, we want to] be able to use … single AI agents that are bespoke with their LLMs for specific problems,” McNamee said. “[We are also looking at] agentic orchestration, where you have a central AI agent that has 30 or 50 different agents that it can call in a microservices way to answer specific questions.”

Enterprise Automation at CIA

During the AWS Public Sector Summit in June, CIA’s Chief Artificial Intelligence Officer Lakshmi Raman described agentic AI as a powerful tool for enterprise automation, capable of executing complex, multi-step workflows and interacting across diverse databases and systems.

“I think AI agents are really exciting to help us in our business use cases,” Raman said. “AI agents can help us with our help desk. AI agents can help us fill out forms automatically so that then we can go look and make sure it’s all been addressed and hit ‘submit.’”

Raman noted that CIA is pursuing responsible deployment of AI agents at the agency. She raised concerns about model “drift,” where AI performance may degrade over time due to shifting data and the “black box” nature of AI systems, which can obscure how decisions are made.

“A lot of things are more normal. What is happening when you thinking about drift, you also have to think about a bigger black box [in AI systems],” Raman said. “What’s going on inside of that black box and can you have a level of explainability around that to your users?”

Raman underscored the importance of maintaining human oversight at critical decision points to ensure accuracy and accountability. Data analysis by CIA staff must always be done “with a human” to review data analysis.

“I think there’s a lot of opportunity there for us again to get the productivity gains that we’re looking for,” she said.

Raman said that that agency needs to maintain rigorous checks to ensure that AI data usage aligns with policy and legal frameworks. Despite these hurdles, she said she is optimistic about agentic AI’s role in enhancing mission support and operational efficiency, provided it is implemented with transparency, oversight and a clear understanding of the technology’s limitations.

“Being able to trust AI has an incredibly important concept. Having the ability to explain that to them is very important for legal and data compliance,” Raman added. “Being able to gather data from multiple slate spaces to be able to leverage AI agents for our cognitive agents, is incredibly exciting.”

Agentic AI for National Security

The Department of Defense is actively exploring agentic AI to enhance military capabilities, moving beyond traditional, pre-programmed systems toward more dynamic and resilient autonomous operations for complex tasks, officials said.

Jamie Fitzgibbon, program manager for the Defense Innovation Unit’s Artificial Intelligence and Machine Learning (AI/ML) said that DOD’s agentic AI evolving from “leader/follower” and “pre-programmed” models to a more intelligent, multi-agent approach. Fitzgibbon used the example of drone swarms to illustrate this change. Instead of a single drone directing others, or all drones being rigidly pre-programmed, agentic AI allows each drone to make individual decisions based on a shared objective.

“[AI agents] all making individual decisions. They’re all programmed with the entire [Concept of Operations] of ‘here’s what we need to do.’ If five of them get taken out, the other five can operate as if all 10 were there, right? They can fill in the gap,” Fitzgibbon explained.

This model is being considered for a wide range of applications, from autonomous submersibles to information operations. Fitzgibbon stressed that agentic AI systems are “hungry” for data and require extensive training, which in turn necessitates new hardware and a forward-thinking policy framework. The goal, she added, is to create models that are not hyper-programmed for a narrow space but are instead capable of intelligent, higher-level decision-making and sensor fusion to provide critical insights to military decision-makers.

For DOD’s software development purposes, agentic AI is becoming more advanced, capable of not only creating code but also deploying it within the entire development environment, according to DOD Chief Software Officer Rob Vietmeyer.

“We’re seeing agentic AI engines, moving not just to developing the software code, but being able to move, being able to deploy into the shell, to deploy right into the full development environment, and not only control the software input, but also starting to control the pipelines,” Vietmeyer told GovCIO Media & Research. “This is fascinating and also really scary to where we’re going on this journey”

Vietmeyer echoed the need for responsible agentic AI across DOD, saying security standards are critical to any national security implementation.

“How do we make sure that these AI agents are operating within a zero-trust framework, that we’re creating hermetic builds, and these AI agents aren’t compromising these controls in some way?” Vietmeyer said.