First-ever public database documenting deployed AI agents developed to address gaps in transparency and safety

Agentic AI systems are defined by their ability to plan and execute tasks autonomously, exhibiting properties such as goal-directedness, long-term planning, and direct action-taking. These systems extend beyond traditional AI models by integrating reasoning frameworks, memory mechanisms, and external tool usage to perform tasks such as software engineering, research assistance, and even computer-based interactions.

Artificial Intelligence is entering a new phase of development with the rise of agentic AI systems - autonomous digital entities capable of executing complex tasks with minimal human intervention. While these advancements bring significant opportunities, they also pose considerable risks that remain inadequately documented.

In a groundbreaking effort to bridge this gap, researchers from leading institutions - including MIT, Stanford, Harvard, and the Hebrew University - have developed "The AI Agent Index," a structured database that systematically records deployed agentic AI systems. Published as a preprint, the study presents crucial insights into the components, applications, and risk management practices of these AI systems.

Understanding agentic AI systems

Agentic AI systems are defined by their ability to plan and execute tasks autonomously, exhibiting properties such as goal-directedness, long-term planning, and direct action-taking. These systems extend beyond traditional AI models by integrating reasoning frameworks, memory mechanisms, and external tool usage to perform tasks such as software engineering, research assistance, and even computer-based interactions. The study highlights how agentic AI is transitioning from experimental prototypes to widely deployed technologies in industries such as cybersecurity, robotics, and scientific research.

Despite their increasing sophistication, agentic AI systems remain largely undocumented, leading to uncertainties regarding their safety and ethical implications. The researchers emphasize that, while developers frequently disclose system capabilities and applications, transparency around risk management and safety measures is lacking. This knowledge gap raises concerns about potential cybersecurity threats, system manipulation, and unforeseen consequences from autonomous decision-making.

The AI Agent Index: A new standard for AI documentation

To address these concerns, the AI Agent Index offers the first structured framework for documenting AI agents comprehensively. It catalogues 67 deployed agentic AI systems, providing detailed insights into their technical components, intended uses, and governance structures. The index documents key elements such as base models, planning and reasoning mechanisms, observation and action spaces, and user interfaces.

Moreover, the index sheds light on safety practices - or the lack thereof. While 70.1% of indexed systems provide some public documentation, only 19.4% disclose a formal safety policy, and fewer than 10% report external safety evaluations. This discrepancy underscores the urgent need for improved safety oversight, transparency, and governance mechanisms as these systems become more integrated into critical sectors.

The study reveals several significant trends within the agentic AI landscape. Firstly, most agentic AI systems are developed by corporate entities rather than academic institutions, with U.S.-based companies leading the field. Industry-driven development raises important questions about corporate incentives, potential monopolization, and proprietary restrictions that may hinder broader transparency efforts.

Additionally, software engineering and computer-based applications dominate the deployment landscape, with 74.6% of indexed AI agents specializing in these domains. This focus on digital automation presents a double-edged sword: while these systems enhance efficiency and productivity, they also introduce novel risks such as code vulnerabilities, biased decision-making, and reduced human oversight in critical operations.

Furthermore, the study highlights the limited regulatory frameworks governing AI agents. With most developers failing to disclose meaningful safety evaluations, policymakers are faced with the challenge of establishing standardized guidelines for AI agent transparency, accountability, and risk mitigation. The authors propose various governance strategies, including structured bug bounty programs, systematic AI risk assessments, and the integration of AI agent indices into broader regulatory frameworks.

The future of AI agent documentation

The AI Agent Index marks a significant step toward responsible AI documentation, yet it also underscores the challenges ahead. As AI agents become more complex and embedded in real-world applications, continuous updates to the index will be essential to track emerging risks and advancements. Future iterations may include expanded safety assessments, standardized agent evaluations, and improved collaboration between AI developers, researchers, and regulators.

Ultimately, the study serves as a call to action for greater transparency in AI development. By systematically documenting AI agent characteristics and safety measures, the AI Agent Index lays the groundwork for a more informed and accountable AI ecosystem. As the field progresses, ensuring that AI systems are developed responsibly will be critical in harnessing their potential while mitigating risks for society at large.