Proof-of-Agentness: Towards Verifying Agents on the Web
Zuyao Xu, Xiang Li
Evidence that a Web client is an AI agent, through active challenges and passive interaction signals.
Why verify agentness?
As Web services begin to admit AI agents, distinguishing agents from simple scripts becomes a different question from distinguishing humans from bots. Proof-of-Agentness studies evidence that a client performs agent-like inference rather than cheaply replaying requests.
Two complementary approaches
The proposed framework combines active challenge-response verification with passive analysis of service-visible interaction signals. Verified agent declarations and an AgentScore provide a design for guiding access to protected operations.
Initial evaluation
Eight language models achieve 96.5% overall accuracy on the paper’s challenges, with an average overhead of 1,647 tokens. An IRB-approved interaction study includes twelve human participants and representative Bot and Agent implementations; four classifiers average 95.8% classification accuracy. These are initial results rather than a universal guarantee of agent identification.