Artificial intelligence (AI) agents—software systems capable of planning, reasoning, and acting with varying degrees of autonomy—are rapidly reshaping how the internet functions. Unlike earlier forms of automation, agents increasingly operate across multiple systems, interact directly with other agents, and execute multi-step tasks without continuous human oversight. This evolution marks a structural shift from an internet primarily mediated by human decision making to one characterized by machine-to-machine interaction.
While agentic systems promise efficiency, scalability, and new forms of economic and social coordination, they also strain foundational assumptions about digital identity and attribution, agency and responsibility, and security. Existing governance models—characterized by human users, discrete transactions, and relatively static software—struggle to address persistent, adaptive, and cross-domain agent behavior.
Given these developments, this white paper raises three core insights into the AI agent landscape. It also presents corresponding recommendations for policymakers, industry leaders, and standards bodies. Ultimately, our work seeks to ensure that AI agents enhance, rather than erode, trust and stability in the digital ecosystem.
Identity and Attribution
Identity and attribution frameworks must evolve to account for persistent, cross-system agents whose actions may not map cleanly onto individual human operators as principals.
- Develop mechanisms to trace agent identities and actions across time and systems, recording provenance, authorization scope, and execution context. This is especially critical in multi-agent interactions to better enable investigators to reconstruct not only what actions occurred, but how authority and decision making propagated through a system.
Evaluation
Evaluation frameworks must evolve to be accurate and reliable in order to cultivate real-world trust. Current benchmarks largely measure progress in technical capabilities, accuracy, and task completion rates, but reliability and trustworthiness in operational environments remain poorly understood. Establishing consistent, deployer-relevant evaluation frameworks is critical to anticipate risk, guide deployment, and earn user confidence.
- Map agent behaviors—such as autonomy level, tool access, memory persistence, and cross-system interaction—to trust impact scores. These scores would reflect the consequence of agent risks and failures on user confidence, system integrity, and institutional trust.
Legal Doctrine for Authority
Legal doctrine must evolve to fully map the spectrum of risks that agents acting under delegated authority can introduce. Traditional principal-agent, product liability, and contract law provide only partial guidance, leaving gaps in accountability when harm emerges from autonomous or semi-autonomous agent actions. Addressing these gaps is essential not only for managing risk and ensuring recourse for affected parties, but also for enabling responsible deployment and scaling of agentic systems.
- Treat authorization as a dynamic, revocable process based on observed behavior, environmental context, and risk signals, to account for the fact that agent behavior evolves over time.
