The Central Intelligence Agency used AI to generate an intelligence report for the first time in its history. That milestone, disclosed by Deputy Director Michael Ellis on Thursday, marks a structural threshold in how the agency will approach its core mission. Within a decade, Ellis predicts, CIA officers won't just work alongside AI tools. They'll manage teams of them.
Speaking at a Special Competitive Studies Project event in Washington, Ellis outlined a three-phase trajectory for the CIA's AI integration. The first phase is already underway: deploying AI "coworkers" into agency analytics platforms to assist with drafting, editing, and comparing reports against tradecraft standards. The second phase, Ellis said, is treating AI as an "autonomous mission partner." The third phase is the hybrid model — officers running coordinated teams of AI agents to scale the speed and volume of intelligence work.
The pace implied in that roadmap is notable. Ellis didn't frame this as a 20-year vision. He said "within a decade."
The Infrastructure Underneath
What's striking is that the CIA has already built significant scaffolding for this transition. The agency is operating more than 300 AI projects. It doubled its technology-related foreign intelligence reporting on how adversaries like China are deploying advanced AI and related technologies. It elevated its Center for Cyber Intelligence into a full mission center — a reorganization Ellis said is already "paying dividends" by enabling faster tool deployment to the field.
Perhaps most concretely, the CIA published a new Acquisition Framework in February. The details matter: a centralized vendor vetting system and streamlined IT authorization process designed to compress the time between identifying a mission requirement and receiving operating authority. Director John Ratcliffe called for "a radical shift towards a culture of speed, agility, and innovation." That language is not standard agency boilerplate. It suggests the acquisition bottlenecks that have historically delayed tech adoption are now understood as a mission risk.
The Vendor Problem Ellis Named
Ellis didn't mention Anthropic by name, but the subtext was unmistakable. "We cannot allow the whims of a single company to constrain our use of AI," he said, adding that the CIA is pursuing vendor diversification "to preserve operational freedom."
This is a direct reference to the escalating dispute over Anthropic's domestic surveillance and autonomous weapons restrictions, which led the Defense Secretary to designate the company's products a "supply chain risk" and triggered a White House directive phasing out Anthropic tools across federal agencies. The CIA's public posture — building toward AI agent teams while explicitly hedging against single-vendor dependency — reads as a deliberate signal that the agency intends to move fast without surrendering leverage.
Ellis did not address Anthropic's Project Glasswing, the consortium announced this week meant to surface software vulnerabilities using frontier AI. The intelligence community and its industry partners are already examining how such capabilities affect cyber missions, Nextgov/FCW reported Wednesday.
What This Means for the Ecosystem
The CIA's trajectory has two implications for the agent infrastructure ecosystem that go beyond the obvious.
First, the agency is not approaching AI as a point solution. The roadmap — from coworker to autonomous partner to agent team manager — describes an evolution in how human analysts relate to AI systems over time. That's a specific kind of software architecture problem: systems that start as tools, develop into collaborative agents, and eventually become managed workloads. The vendors who can support that full arc, rather than just the current phase, are the ones the CIA will need.
Second, the acquisition framework matters for vendors who aren't household names. A centralized vendor vetting system with a streamlined authorization process lowers the barrier for smaller AI providers to enter the government market. If the CIA follows through on Ratcliffe's language about speed and agility, the procurement pathway for novel agent frameworks could look meaningfully different than it did two years ago.
The first AI-generated intelligence report is already done. The infrastructure to scale that capability is being built now. The question for the vendor ecosystem is whether the CIA's stated commitment to speed and diversification will survive contact with the institutional realities of government procurement — and which companies are positioned to be part of the next architecture.
Primary sources: Defense One — CIA employees will get AI 'coworkers'; CIA.gov — Acquisition Framework announcement; POLITICO — CIA is trusting AI to help analyze intel; Nextgov — CIA plans for 'AI coworkers'; Studies in Intelligence, Vol. 70, No. 1 (CIA)