Agentic artificial intelligence (AI) represents a paradigm shift from current AI systems that merely respond to prompts to autonomous entities capable of setting goals, creating plans, and executing multi-step tasks with minimal human oversight. According to experts at the Special Competitive Studies Project (SCSP), a nonprofit and nonpartisan initiative focused on strengthening America's long-term competitiveness in AI, this evolution marks a critical juncture where "AI is beginning to help build better AI," potentially leading to a self-accelerating loop that compounds capability development far beyond current projections.
Ylli Bajraktari, president of SCSP, emphasized in a recent newsletter that agentic AI systems could qualitatively expand adversarial capabilities. "An agent that can navigate complex bureaucratic systems, identify exploitable vulnerabilities, and act without leaving a clear attribution trail represents a qualitative expansion of adversarial capability," he stated. This poses significant risks for global security, as adversaries may deploy such systems in areas with weak governance for coercion, espionage, and influence operations.
Contrary to common policy assumptions, effective governance of agentic AI does not hinge on controlling the AI model itself but rather on the scaffolding built around it. SCSP experts outlined key components of this scaffolding: connectors to bridge the model to real-world infrastructure like email and financial platforms; memory for learning and adaptation over time; planning capabilities to break down large objectives into smaller tasks; permission structures defining system access; and guardrails determining what the system will refuse to do, such as spending limits or human sign-offs.
Accountability remains a major challenge, and current governance frameworks are falling short in three critical ways. First, responsibility becomes untraceable when AI agents act autonomously, making it impossible to determine who authorized specific actions. Second, existing frameworks focus on task completion rather than evaluating whether the AI performed tasks safely or caused harm. Third, agentic AI builds detailed personal profiles by accumulating data on behavior patterns, preferences, and inferences, often capturing more sensitive information than individuals intend to share.
Despite these challenges, SCSP experts caution against fearing or deferring agentic AI. Instead, institutions that prioritize understanding, shaping, and governing this technology will determine their competitive position and influence the global operating environment for agentic AI. To learn more about how the United States should pursue effective governance of agentic AI, visit scsp.ai.
