Agentic AI Evaluation

No single definition of agentic AI has gained universal acceptance, which can make it harder to ground measurement and evaluation frameworks, and the policy decisions that depend on them. Researchers, developers, and policymakers have proposed their own definitions, often emphasizing different features and trade-offs. “Agency” spans a spectrum across several practical dimensions, which may include autonomy, planning horizons, tool use and environment coupling, and adaptivity. Familiarity with this spectrum may inform both technical evaluation and policy considerations related to evaluation, accountability, and deployment contexts.

While AI evaluation for conventional AI tools is already subject to several well-documented limitations, evaluating agentic AI systems presents a considerably greater technical challenge. Core capabilities remain difficult to measure systematically; autonomy levels vary widely across implementations; social intelligence can manifest differently in controlled settings versus real-world environments; and long-horizon planning complicates standardization through conventional benchmarks. Current evaluation frameworks tend to focus narrowly on capability and accuracy, may not fully reflect real-world deployment conditions, and operate without agreed-upon standards for scientific validity.

The measurement gap has practical consequences for deployers making adoption decisions. Organizations may find it difficult to deploy systems whose behavior they cannot reliably predict, and without reliable evaluation frameworks, it can be challenging to fully assess risks around security, liability, misuse, and systemic harms. This challenge is even more pronounced where agentic AI systems are acting on behalf of, rather than in conjunction with, human operators. In many deployment scenarios, users are expected to delegate their workflows, private information, or online identities to these systems, and difficulty anticipating performance or failure modes may create that slows broader deployment and integration of agentic AI into business-critical workflows. As agentic AI systems are integrated into increasingly sensitive domains and become more widespread across our economy, expanding coverage, reproducibility, and comparability will be important to help evaluations remain transparent and can provide insights that are useful to consumers, deployers, and policymakers.

• Develop a collaborative community spanning academia, industry, government, and civil society.
• Bring together stakeholders who understand how measurement and evaluation drive innovation in agentic AI design, policy, and governance.

• Address measurement and evaluation challenges along the entire agentic AI stack–from deployment to workflow integration.
• Create trustworthy telemetry and instrumentation designed to provide data at different levels of abstraction (e.g., scratch pad COT, tool use and invocation logs, cognitive logs using models such as belief, desire and intention) to support auditable, interpretable and governable agentic AI models and systems.
• Extend beyond basic evaluation to create frameworks for red-teaming, blue-teaming, and field testing that measure both immediate and long-term benefits and risks of agentic AI deployment.

• Identify and pursue opportunities for joint initiatives between academia, industry, civil society, and government.
• Target key deliverables, including standardized benchmarks, open-source evaluation tools, and technical sandboxes that can support evidence-based policy development.

• Develop a roadmap for agentic AI governance grounded in measurement and evaluation practices. Leverage the expert network to both craft this framework and use it to inform policymakers at state, federal, and international levels.