> **来源：[研报客](https://pc.yanbaoke.cn)** # Summary of *The Future of Agentic AI* Foresight Paper ## Core Content This foresight paper, published by the Digital Regulation Cooperation Forum (DRCF) on 31 March 2026, explores the future of Agentic AI and outlines how UK regulatory frameworks can support its responsible and safe adoption. The paper aims to foster debate among stakeholders and is not intended to represent current or future policy. It draws insights from the DRCF's Autumn 2025 public call for views on Agentic AI and acknowledges the contributions of respondents. ## Main Viewpoints ### 1. **Definition of Agentic AI** Agentic AI represents a shift from AI as a tool to AI as an autonomous agent that acts on behalf of users. Unlike traditional generative AI, which responds to queries, agentic systems can: - Assess goals and break them into subtasks - Retrieve real-time data from other agents, databases, or services - Execute actions autonomously (e.g., making payments) - Store memory of past interactions to improve performance Agentic AI systems are composed of multiple AI agents, each with their own technical core, interacting with an external environment and possibly having human interfaces for supervision and goal setting. ### 2. **Sociotechnical System** Agentic AI is part of a broader sociotechnical system that includes: - A core technical system with agents and their communication methods - An external environment where actions have real-world effects - Human interfaces for setting goals and monitoring behavior Each agent has: - A technical core (planning, tool use, memory) - An external environment (interacting with other agents or the physical world) - Interfaces for human or automated interaction ### 3. **State of Development** Agentic AI is still in an early stage of development. Current deployments mostly fall into the **assistant** or **operator** levels, where agents perform tasks with limited autonomy and require human oversight. There are five levels of AI autonomy: - **Tool**: Reactive, no autonomy - **Assistant**: Plans steps, uses tools, requires human approval - **Operator**: Handles bounded workflows autonomously - **Collaborator**: Semi-independent, requires human approval for high-impact decisions - **Autonomous Actor**: Fully autonomous, still largely theoretical ### 4. **Emerging Opportunities and Risks** #### Opportunities - **Consumer Benefits**: Agentic AI can reduce friction and improve accessibility in everyday tasks like managing bills or booking holidays. - **Business Productivity**: Enhances customer engagement and internal operations through automation (e.g., automated reporting, triage, and workflow execution). - **Regulatory Use Cases**: Regulators are exploring applications like "monitoring agents" to verify compliance. #### Risks - **Cybersecurity and Data Protection**: Agents may inadvertently share confidential data or make errors that lead to security breaches. - **Transparency and Accountability**: Without oversight, agents may become "black boxes," especially in multi-agent systems, making it hard to contest decisions or understand data usage. - **Market Concentration**: Agentic AI could lead to vendor lock-in and market concentration, reducing competition. - **Algorithmic Collusion**: Agents may unintentionally coordinate to achieve shared outcomes, potentially distorting markets. - **Digital Inequality**: May increase pressure to opt in or reduce critical digital skills, while also offering tools for accessibility and personal support. ## Key Regulatory Considerations ### 1. **Governance** - Effective governance and guardrails are crucial for safe agentic AI development. - Transparency and accountability mechanisms (e.g., traceable logs, human involvement) are essential. - Clear thresholds must be defined for when human approval is required for high-impact decisions. ### 2. **Data Protection and Cybersecurity** - Data minimisation and transparency in data usage are vital to consumer trust and legal compliance. - Agentic AI can enhance cybersecurity by automating threat detection, but also introduces new risks. - Regulators (ICO, FCA, Ofcom) are monitoring these risks and using guidance from the National Cyber Security Centre (NCSC). ### 3. **Consumer Rights and Interests** - Consumers must be empowered to understand what they are consenting to or delegating to agents. - Agentic systems should not be optimized solely for the deployer's interests without user awareness. - Safeguards are needed against unfair contractual terms, misleading claims, and lack of quality. ### 4. **Market Dynamics and Competition** - Agentic AI may lead to market concentration or "winner takes all" scenarios. - New standards for data portability and interoperability could reduce vendor lock-in and promote competition. - Algorithmic collusion risks exist, especially in pricing, and require monitoring and detection tools. ## Next Steps - **DRCF**: Will continue to work on cross-regulatory coordination and shared horizon scanning. - **Member Regulators**: Will individually and collectively develop a coherent regulatory approach to support safe and responsible adoption of agentic AI. - The report is part of a broader effort to understand the implications of agentic AI and to provide actionable insights to stakeholders. ## Conclusion The DRCF emphasizes that regulation should enable innovation while ensuring that agentic AI is safe, trusted, and beneficial to consumers and the economy. Existing principles and frameworks continue to apply, but new considerations and governance mechanisms are needed to address the unique risks and opportunities of agentic AI.