Key takeaways

  • Unlike other AI systems that mainly analyze or recommend, agentic AI can pursue goals, decide what steps to take and act with less human input.
  • Agentic AI can amplify familiar AI risks because errors, poor instructions or manipulated inputs can trigger chains of actions before humans can intervene.
  • Governance should adapt existing rules and safeguards to agentic AI, while preparing for interactions between AI agents, including those operated by different actors.

Agentic AI systems can pursue objectives over multiple steps, determine intermediate actions, use tools, adapt to changing circumstances and act with reduced human intervention. This publication examines what these capabilities could mean for international peace and security. It distinguishes agency from autonomy and assesses how agentic capabilities are emerging across cybersecurity, military command and control, and autonomous physical systems. The evidence remains uneven: the clearest current examples are found in cyber and other digital environments, while military applications in command and control and physical systems remain more experimental.

The publication also examines risks including misalignment, error propagation, constraint circumvention, loss of human control and interactions between agents operated by different actors. For international peace and security, these developments could compress decision time and reduce opportunities for human clarification and de-escalation, particularly as agents operated by different actors increasingly interact. This is especially important where existing crisis-management arrangements rely on human signalling, interpretation and restraint.

Rather than assuming that agentic AI requires an entirely new governance regime, the publication argues for stress-testing existing legal and policy frameworks, strengthening practical controls and assurance, and beginning international dialogue on agent-to-agent interactions before such practices become widespread.