CIA Says Agentic AI Can Turn Weeks of Intelligence Work Into Hours

By Olivia Hart

Agentic tools can compress intelligence work that once took CIA officers weeks into hours, two CIA leaders said at ClearanceJobs Connect 2026 in Reston, Virginia.

On a panel with AWS partner Amber Nicholas, Larry and Chris described the agency’s move from chatting with artificial intelligence to handing it tasks. The CIA now refers to AI as “SI,” and the panel set out what that shift means for intelligence work and the cleared workforce.

Larry is moving into a new role as dean of his directorate’s university. For this shift, he prefers to measure by “value of investment,” meaning time returned to the mission, rather than by dollars saved. That makes time, not a headline savings target, the program’s operating measure.

From chatting to task handoff

The change is practical: work that once consumed weeks can now finish in hours. Officers will hand more assignments to agents, but the agency is not treating those systems as independent decision-makers. Human ownership remains the central rule.

Chris, chief scientist in the CIA’s Office of Super Intelligence, said: “The biggest barrier isn’t computing power or better models. It’s basic AI literacy.” He said the strongest signal that a candidate will succeed is that the person already uses these tools in daily life.

That signal can show up in several forms:

• A GitHub portfolio still matters.
• So does building a scheduling app for your kids’ ball games.
• Daily use of the tools matters more than a résumé claim of agentic skill.

Accountability before scale

Both speakers said a human must own every outcome. The agency uses an AI governance board, brings legal into review, and trains analysts never to assume AI output is fully accurate.

As agents multiply, identity and observability become essential. Every agent action has to trace to a person or system and to a purpose. The agency therefore needs a clear chain of responsibility, not simply access to more capable models.

The controls preserve the analyst’s role. Agents can carry out tasks, but officers remain accountable for the result and must validate outputs instead of accepting them as automatic truth.

Hiring meets evidence

Chris noted a surge in demand for cleared forward deployed engineers. Employers also need evidence that claimed agentic coding skills produce strong code, rather than treating the label as proof.

Larry warned that some new graduates claim agentic coding skills but produce weak code. He recommended that employers validate those skills directly. He even floated switching the models off one day a week, making developers demonstrate that they can still code on their own without the tools.

The proposal was a check on professional skills: if AI handles part of the workflow every day, developers should retain the underlying ability to build without it.

The classified-environment requirement

AWS’s Amber Nicholas said vendors need to understand classified environments, where a tool that requires an internet connection simply will not work. Connectivity is therefore a basic deployment constraint for any vendor serving the agency.

What does this change for cleared professionals?

It changes the evidence employers should look for and the habits they should cultivate:

• Use AI tools in daily work, not only in demonstrations.
• Validate coding ability with real work rather than accepting an “agentic” label.
• Know how every agent action traces back to a person or system and a purpose.
• Build the judgment to challenge output that may be inaccurate.

For the CIA, the near-term priority is people who can work with agents, prove their skills and remain accountable for every outcome—the panel’s simple instruction was to get hands-on and use AI to teach yourself AI.