GPAI - Autonomous AI Framework

Overview

The Global Partnership on Artificial Intelligence (GPAI) is a leading multi-stakeholder initiative dedicated to guiding the responsible development and use of artificial intelligence. In line with its mission, the GPAI Agentic AI Systems Framework (gpai-agentic-ai-framework-2025) represents a significant undertaking to address the unique opportunities and challenges presented by agentic AI systems. These advanced AI systems are characterized by their capacity for autonomous decision-making and goal-oriented action with limited human oversight, marking a new frontier in AI capabilities. The framework, anticipated to be a key output from a project launched in 2025, aims to provide comprehensive guidance for developers, deployers, and users of such systems, ensuring their integration into society is both beneficial and aligned with human-centric values. The initiative underscores GPAI’s commitment to bridging the gap between theoretical AI principles and practical implementation, fostering international cooperation to establish common understandings and best practices for this rapidly evolving technology.

The development of the GPAI Agentic AI Systems Framework is particularly pertinent given the increasing sophistication and autonomy of AI. Agentic AI systems, by their very nature, introduce new layers of complexity in terms of governance, risk management, and ethical considerations. The framework is expected to build upon the foundational OECD AI Principles, which serve as the guiding ethical framework for GPAI's work, translating these high-level principles into actionable recommendations specific to agentic capabilities. This proactive approach by GPAI, particularly through the leadership of expert support centers like the Tokyo Expert Support Center, highlights the urgency in establishing robust governance mechanisms for AI that can operate with increasing independence. The framework will likely address critical areas such as ensuring transparency, accountability, safety, and the effective balance of human control with AI autonomy, thereby fostering trust in these powerful new systems.

Definitions

The GPAI Agentic AI Systems Framework is expected to establish clear definitions to delineate the scope and nature of agentic AI. Central to this framework is the concept of 'Agentic AI Systems,' which refers to AI systems capable of autonomous decision-making and goal-oriented actions with limited direct human intervention. These systems are designed to pursue complex objectives adaptably, employing reasoning and planning capabilities to execute workflows, utilize tools, and interact with both other AI agents and human users. Unlike simpler AI applications that primarily generate content or answer questions, agentic AI systems exhibit a higher degree of independence, learning, and adaptation over time to achieve predefined goals. This distinction is crucial for identifying the specific governance and risk management challenges associated with their deployment.

Furthermore, the framework will likely differentiate Agentic AI Systems from 'General-Purpose AI (GPAI) Models,' a term also gaining prominence in international AI governance, notably within the EU AI Act. General-Purpose AI models, often exemplified by large language models, are foundational models capable of competently performing a wide array of distinct tasks. While these models can be integral components of Agentic AI Systems, they do not, on their own, constitute a complete agentic system. The framework is expected to clarify the interplay between these foundational models and the broader agentic systems built upon them, addressing the distinct responsibilities of model providers versus system deployers. Understanding these definitional nuances is fundamental for applying appropriate regulatory and ethical safeguards, ensuring that the framework provides clear guidance across the entire AI value chain for agentic technologies.

Governance and Institutional Framework

The GPAI Agentic AI Systems Framework is developed within the robust governance structure of the Global Partnership on Artificial Intelligence, an integrated partnership that merged with the OECD in July 2024. This integration leverages the synergies between GPAI's multi-stakeholder expertise and the OECD's established work programme on AI, ensuring a comprehensive and inclusive approach to AI governance. GPAI's mission, supported by a diverse expert community comprising governments, industry, academia, and civil society, is to advance human-centric and trustworthy AI, grounded in the OECD AI Principles. The development of frameworks like the Agentic AI Systems Framework is typically spearheaded by GPAI's expert Working Groups, particularly the Working Group on Responsible AI, which is mandated to foster the responsible development, use, and governance of human-centred AI systems. This collaborative model ensures that the framework benefits from a wide array of perspectives and technical insights.

A critical component of GPAI's institutional framework for developing such instruments is its network of Expert Support Centers located in Montreal, Paris, and Tokyo. These centers play a pivotal role in shaping and implementing GPAI's work plan through practical projects and applied research. Notably, the Tokyo Expert Support Center was explicitly "set to lead a project on Agentic AI in 2025," indicating its central role in the conceptualization and development of the GPAI Agentic AI Systems Framework. This distributed yet coordinated structure allows GPAI to address emerging AI challenges, such as those posed by agentic systems, by combining cutting-edge research with practical policy development. The framework, once finalized, will be a testament to GPAI's capacity to translate high-level ethical guidelines into concrete, actionable governance instruments, promoting international cooperation and alignment on responsible AI innovation.

Key Provisions

While the GPAI Agentic AI Systems Framework is still in its developmental phase, its key provisions are expected to address the critical challenges and opportunities presented by autonomous AI. Drawing from existing discussions and the broader responsible AI mandate, the framework will likely emphasize robust risk identification and assessment. This would involve defining and classifying various AI-related risks, particularly those stemming from an agent's autonomy, potential for unintended consequences, and systemic harms. Provisions would likely mandate comprehensive impact assessments throughout the AI system's lifecycle, from design to deployment, to proactively identify and mitigate risks such as labor displacement, incorrect or undesired actions, and security vulnerabilities.

Another central pillar of the framework will undoubtedly be transparency and accountability. Given the independent decision-making capabilities of agentic AI, the framework is expected to propose mechanisms for ensuring that the operation, decision-making processes, and outcomes of these systems are understandable and explainable to human users and oversight bodies. This could include requirements for clear disclosure when AI is being used, logging of agent actions, and mechanisms for auditing and tracing decisions back to their origins. Furthermore, the framework will likely address the importance of human oversight and control, defining thresholds for autonomy and outlining scenarios where human intervention is necessary or mandated. Ethical considerations, such as fairness, non-discrimination, and the protection of fundamental rights, will be woven throughout the provisions, guiding the responsible design and deployment of agentic AI systems to ensure they serve the public good.

Scope and Application

The GPAI Agentic AI Systems Framework is designed to have a broad yet targeted scope, aiming to provide guidance that is applicable across various sectors and stakeholders involved in the lifecycle of agentic AI systems. It is expected to apply to developers who design and build these autonomous systems, deployers who integrate them into specific applications and environments, and users who interact with or rely on agentic AI for various tasks. The framework's applicability will likely extend to both private and public sector entities, recognizing that agentic AI has transformative potential and associated risks across all domains, from enterprise operations to public services. The goal is to establish a common understanding and set of expectations for responsible development and use, regardless of the specific application or industry.

Geographically, as an initiative of the Global Partnership on Artificial Intelligence, the framework is intended to foster international alignment and serve as a reference point for countries worldwide. GPAI's membership, encompassing 44 countries, ensures that the framework's principles and recommendations are developed with a global perspective, promoting interoperability and reducing the potential for divergent regulatory approaches. While not a legally binding treaty, its influence is expected to be significant, guiding national AI strategies, informing the development of domestic regulations, and shaping industry best practices. The framework will likely emphasize scalability and adaptability, acknowledging the rapid evolution of AI technology and the need for guidance that can remain relevant amidst continuous innovation. It aims to be a living document, subject to review and updates as agentic AI capabilities mature and societal impacts become clearer.

Implementation Framework

The implementation of the GPAI Agentic AI Systems Framework will be guided by GPAI's practical, action-oriented approach, aiming to translate high-level principles into tangible tools and methodologies. GPAI's core mission involves supporting cutting-edge research and applied activities to address real-world policy challenges, and this framework will be no exception. It is expected to propose a multi-faceted implementation strategy, including the development of practical toolkits, guidelines, and best practices that organizations can adopt to ensure responsible development and deployment of agentic AI. These resources would likely cover areas such as technical specifications for safety and reliability, ethical impact assessment methodologies, and mechanisms for auditing and validation of autonomous systems. The framework will aim to provide concrete steps for integrating responsible AI practices into existing organizational governance structures.

Furthermore, the implementation framework is likely to emphasize capacity building and knowledge sharing among stakeholders. This could involve workshops, training programs, and expert forums facilitated by GPAI's Expert Support Centers, enabling practitioners to understand and apply the framework's recommendations effectively. The collaboration among GPAI's various Working Groups, particularly Responsible AI and Data Governance, will be crucial in developing holistic implementation guidance that considers the interconnected aspects of agentic AI, from algorithmic transparency to data privacy. By offering practical, tested frameworks and fostering a collaborative environment, GPAI intends to empower governments, industry, academia, and civil society to proactively manage the risks and maximize the benefits of agentic AI, ensuring its development aligns with the shared values of human rights and democratic principles.

Monitoring and Evaluation

The GPAI Agentic AI Systems Framework will incorporate mechanisms for continuous monitoring and evaluation, reflecting GPAI's commitment to adaptive governance in a rapidly evolving technological landscape. Given the dynamic nature of agentic AI, effective oversight requires ongoing assessment of both the systems themselves and the efficacy of the governance frameworks in place. GPAI, in its integrated partnership with the OECD, regularly monitors global AI policy developments and the implementation of AI principles across member countries. This broader monitoring capability will provide a crucial backdrop for evaluating the impact and relevance of the Agentic AI Systems Framework. The framework is expected to recommend specific metrics and indicators for assessing the performance, safety, and ethical compliance of agentic AI systems in real-world applications, encouraging organizations to adopt robust internal monitoring practices.

Evaluation of the framework's effectiveness will likely involve periodic reviews by GPAI's expert community and ministerial bodies. These reviews would assess whether the framework's provisions are adequately addressing emerging risks, fostering responsible innovation, and promoting international alignment. Feedback mechanisms, potentially including reporting requirements for organizations deploying agentic AI, could be integrated to gather empirical data on implementation challenges and successes. The goal is to ensure that the framework remains relevant and responsive to technological advancements and societal impacts. By fostering a culture of continuous learning and adaptation, GPAI aims to ensure that its guidance on agentic AI systems not only sets initial standards but also evolves to meet future challenges, thereby maintaining trust and promoting the beneficial development of these powerful technologies.

Relationship to Other Instruments

The GPAI Agentic AI Systems Framework is intrinsically linked to and builds upon a broader ecosystem of international AI governance instruments, most notably the OECD Recommendation on Artificial Intelligence (OECD AI Principles). These principles, adopted in 2019 and updated in 2024, serve as the foundational ethical and policy framework for all of GPAI's work, providing the common ground for human-centric, safe, secure, and trustworthy AI. The Agentic AI Framework is designed to operationalize these high-level principles specifically for the unique characteristics of autonomous agentic systems, translating theoretical commitments into practical guidance. This relationship ensures coherence and avoids fragmentation in the global AI governance landscape, reinforcing a shared vision for responsible AI development.

Furthermore, the framework will likely acknowledge and interact with other significant international and regional initiatives. For instance, the EU AI Act, which entered phased implementation with obligations for General-Purpose AI (GPAI) models taking effect in August 2025, represents a major regulatory effort that defines and regulates AI systems and GPAI models. While the GPAI framework is non-binding, it can serve as a complementary instrument, offering detailed best practices and technical guidance that can help organizations comply with evolving regulatory requirements like those in the EU AI Act. Similarly, it will draw connections to other efforts such as the G7 Hiroshima AI Process, which has developed guiding principles and codes of conduct for advanced AI systems, and the Council of Europe's Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law. By situating itself within this broader landscape, the GPAI Agentic AI Systems Framework aims to contribute to global interoperability and a harmonized approach to AI governance.

International Alignment

International alignment is a core tenet of the Global Partnership on Artificial Intelligence, and the Agentic AI Systems Framework is being developed with this objective squarely in mind. GPAI was founded to bridge the gap between theory and practice in AI by fostering international collaboration among leading experts from diverse sectors. The framework's development, led by the Tokyo Expert Support Center's project on Agentic AI in 2025, is a direct manifestation of this commitment to global cooperation. By bringing together insights from 44 member countries and a multi-stakeholder expert community, GPAI aims to create guidance that is globally relevant and adaptable to different national contexts, thereby reducing the risk of regulatory fragmentation and promoting a common understanding of responsible agentic AI.

The framework's alignment efforts extend to its foundational adherence to the OECD AI Principles, which are widely recognized as the first intergovernmental standard on AI and form the basis for many national and regional AI strategies. By building upon these shared principles, the GPAI Agentic AI Systems Framework contributes to a consistent global narrative on AI ethics and governance. It also seeks to complement and inform other international dialogues and initiatives, such as those arising from G7 summits (e.g., the Hiroshima AI Process) and discussions within the UN and Council of Europe, which also address the governance of advanced AI. Through this concerted effort, GPAI aims to ensure that the rapid advancements in agentic AI are managed cooperatively on a global scale, fostering an environment where innovation thrives responsibly and the benefits of AI are shared broadly while mitigating potential harms through harmonized approaches.

Implementation Timeline

Adoption and Endorsement

Sources and References