Introducing Agent Tokens

As AI agents become increasingly sophisticated, the current approach of relying on CAPTCHAs to differentiate between human and AI-generated content on the Internet may prove insufficient. This paper proposes a solution called "agent tokens" to provide a dedicated interface for AI agents to interact with websites. By extending HTTP, agent tokens allow AI agents to identify themselves and provide structured information about their capabilities, intentions, and provenance. This approach encourages cooperation between websites and agent creators, enabling the labeling and handling of AI-generated content separately from human-generated content. The paper outlines the technical foundations of agent Tokens and recent research on agent visibility. It also discusses the potential content of agent tokens, including basic information, structure, licensing, training and evaluation details, intentions, and reach. The limitations of the current proposal, such as the need for independent verification, watermarking, and consensus on design, are acknowledged. The paper concludes by emphasizing the importance of providing a channel for AI agents to co-exist with humans on the Internet while recognizing the need for further research and infrastructure development to ensure the safe integration of this paradigm shift.

Read the full piece here.