Workshop Report: Preparing for AI Agents
Executive Summary
The concept of artificial intelligence (AI) systems that actively pursue goals—referred to as "AI agents"—has been around for some time. However, recent advancements in large language models (LLMs), such as those powering ChatGPT, have sparked renewed interest and excitement among AI developers. Startups and major tech companies are announcing plans to build and deploy sophisticated AI agents that can serve as personal assistants, virtual employees, and more.
These AI agents are characterized by their ability to pursue complex goals in complex environments, exhibiting independent planning and adaptation. For instance, a cyber-offense agent that can autonomously carry out a cyber intrusion would be considered more agentic than a chatbot advising a human hacker. Similarly, a "CEO-AI" that can run a company without human intervention would be more agentic than an AI acting as a personal assistant.
General-purpose LLM-based agents are currently a significant focus of interest among AI developers and investors. These agents consist of advanced LLMs interfaced with external environments and tools like browsers or code interpreters. Early proof-of-concept products that can write code, order food deliveries, and manage customer relationships are already available.
While AI agents offer numerous potential benefits, they also pose significant risks and challenges. The ability of these agents to pursue complex goals without human intervention could lead to more serious accidents, misuse by scammers and cybercriminals, and new challenges in allocating responsibility when harms occur. Data governance and privacy issues may be exacerbated as developers seek to tailor agents to specific users or contexts. Additionally, widespread use of highly capable agents could result in skill fade and dependency among users, as well as labor impacts as an increasing range of tasks become automated.
To address these challenges, the workshop participants discussed three categories of interventions:
- Measurement and Evaluation: Developing better methodologies to track the capabilities and real-world impacts of AI agents and collecting ecological data to anticipate and adapt to future progress.
- Technical Guardrails: Designing AI agents and technical ecosystems to support governance objectives like visibility, control, trustworthiness, security, and privacy. However, there may be trade-offs between different objectives.
- Legal Guardrails: Analyzing existing legal doctrines and frameworks to manage the impacts of AI agents, including questions about the "state of mind" of AI agents, legal personhood, and how existing principal-agent frameworks should apply in situations involving AI agents.
Given the significant interest and investment in AI agents, policymakers should understand the potential implications and intervention points. Steps include improving measurement and evaluation, considering how technical guardrails can support multiple governance objectives, and analyzing how existing legal doctrines may need to be adjusted or updated to handle more sophisticated AI agents.
Key Findings
- Characteristics of AI Agents: Pursue complex goals in complex environments, exhibit independent planning and adaptation.
- Current State: General-purpose LLM-based agents are gaining interest; early products are available.
- Risks and Challenges: Potential for serious accidents, misuse, allocation of responsibility, data governance issues, skill fade, dependency, and labor impacts.
- Intervention Categories:
- Measurement and Evaluation: Develop methodologies to track capabilities and impacts.
- Technical Guardrails: Design technical ecosystems supporting governance objectives.
- Legal Guardrails: Analyze existing legal doctrines and frameworks.
Authors
- Helen Toner, John Bansemer, Kyle Crichton, Matt Burtell, Thomas Woodside, Anat Lior, Andrew J. Lohn, Ashwin Acharya, Beba Cibralic
Acknowledgements
Acknowledgments section omitted for brevity.
Endnotes
Endnotes section omitted for brevity.