2025AI智能体的实践应用_评估与治理基础框架白皮书_34页_7mb

AI Agents in Action: Foundations for Evaluation and Governance

Core Content

This white paper provides a comprehensive overview of the technical foundations, classification, evaluation, and governance of AI agents. It is designed to guide organizations in the responsible adoption and deployment of AI agents, emphasizing the need for structured approaches to ensure safety, trust, and accountability.

Main Points

1. Evolving Technical Foundations of AI Agents

• Software Architecture: AI agents are built with a layered architecture consisting of application, orchestration, and reasoning layers. These layers enable intelligent, context-aware automation and are essential for understanding how agents operate within organizational systems.
• Communication Protocols: Protocols like the Model Context Protocol (MCP) and Agent-to-Agent (A2A) protocol facilitate seamless integration and interoperability between agents and external systems. MCP standardizes connections to enterprise resources, while A2A enables agents to discover, interact, and collaborate with each other.
• Cybersecurity Considerations: As AI agents interact with external tools and systems, they introduce new cybersecurity risks. Security strategies must evolve from perimeter-based models to more advanced, zero-trust approaches. Identity management, micro-segmentation, and ongoing verification are critical for mitigating these risks.

2. Foundations for AI Agent Evaluation and Governance

• Classification: Agents are classified based on their function, role, predictability, autonomy, and authority. These dimensions help define an agent's operational profile and guide its evaluation and governance.
• Evaluation: Evaluation focuses on assessing an agent's performance and limitations in representative environments. It ensures that agents meet expected standards and are reliable in their tasks.
• Risk Assessment: Risk assessment involves analyzing potential harms associated with an agent's actions. It uses classification and evaluation as inputs to identify and mitigate risks effectively.
• Governance: Governance translates evaluation and risk assessment results into proportionate safeguards and accountability mechanisms. It ensures that agents are used responsibly and that their impact is managed appropriately.

3. Looking Ahead: Multi-Agent Ecosystems

• The paper emphasizes the importance of a progressive governance approach to support the development of multi-agent ecosystems. These ecosystems involve multiple agents working together, which requires robust frameworks for coordination, trust, and accountability.
• Early adopters are encouraged to start small, iterate carefully, and apply proportionate safeguards to ensure successful deployment and long-term trust in AI agents.

Key Information

• Adoption Trends: 82% of organizations plan to integrate AI agents within the next one to three years, indicating that most efforts are still in the planning or pilot phase.
• Challenges: AI agents introduce new risks such as goal misalignment, behavioural drift, and emergent coordination failures. Traditional governance models are insufficient, and new approaches are needed to manage these risks.
• Protocols: MCP and A2A are key protocols that enable interoperability and seamless integration of AI agents. AP2 is another emerging protocol that addresses secure financial transactions.
• Governance Approach: A progressive governance model is suggested, which links evaluation and safeguards to an agent's task scope and deployment environment. This ensures that governance is proportionate and effective.

Conclusion

The white paper outlines a structured framework for the responsible adoption and deployment of AI agents. It emphasizes the need for a clear understanding of the technical foundations, classification, evaluation, and governance of AI agents. By adopting a progressive and collaborative approach, organizations can ensure that AI agents are used safely, effectively, and in alignment with human values and needs.