⚡ Quick Summary
Published by the World Economic Forum in collaboration with Capgemini, this white paper sets out foundations for evaluating and governing AI agents, aimed mainly at adopters and their decision-makers. It states three contributions: the technical foundations of AI agents, a functional classification differentiating agents by role, autonomy, authority, predictability and operational context, and a progressive governance approach connecting evaluation and safeguards to an agent's task scope and deployment environment.
The paper is organised around four pillars — classification, evaluation, risk assessment and governance. Classification uses seven dimensions, from function and role to predictability, autonomy, authority, use case and environment. Evaluation covers emerging agent benchmarks such as AgentBench, SWE-bench and HCAST and metrics including task success rate, completion time, error types, tool call success and robustness against edge cases. Risk assessment is presented as a five-step life cycle. Governance scales from a baseline of mechanisms such as least-privilege access control, audit logging and human-in-the-loop or human-on-the-loop oversight. Four case studies — robot vacuum cleaner, coding co-pilot, autonomous vehicle and personal assistant — illustrate the pillars, and a closing section examines multi-agent ecosystems.
🧩 What’s Covered
The paper runs from technical foundations to governance practice.
- Foreword and executive summary: leaders from the World Economic Forum and Capgemini present the paper as guidance for adopters, noting limited existing guidance on designing, testing and overseeing agents responsibly.
- Section 1, technical foundations: four paradigms (classical software, neural networks, foundation models, autonomous control), a three-layer architecture of application, orchestration and reasoning, and protocols MCP, A2A with agent cards, AP2 and the AGNTCY agent connect protocol.
- Cybersecurity considerations (1.3): the shift from perimeter defences to zero trust, the larger attack surface created by MCP and A2A, prompt injection and misuse, and the need for audit trails and incident response.
- Classification (2.1): seven dimensions — function, role, predictability, autonomy, authority, use case, environment — with autonomy likened to SAE International's driving-automation levels, and a robot vacuum cleaner case study.
- Evaluation (2.2): definitions and emerging benchmarks (AgentBench, SWE-bench, HCAST), task- and system-level metrics, three principles (contextualization, multidimensional assessment, temporal and behavioural monitoring), and a coding co-pilot case study.
- Risk assessment (2.3): a five-step life cycle — define context, identify, analyse, evaluate, manage — tabulated with objectives, example activities and outputs, plus inherent versus residual risk and an autonomous vehicle case study.
- Governance (2.4): progressive governance levels; a baseline of mechanisms including least-privilege access control, DPIA and GDPR or CCPA checks, logging, traceability, explainability and manual redundancy; HITL and HOTL oversight; and a personal assistant case study.
- Looking ahead and conclusion (Section 3): multi-agent ecosystems, agent-to-agent commerce, an "internet of agents", governor or auditor agents, embodied agents, and failure modes such as orchestration drift, semantic misalignment and cascading effects.
💡 Why it matters?
The paper addresses a practical gap: organisations are moving agents into workflows while evaluation and oversight practices remain nascent. It gives adopters and providers a shared vocabulary — classification dimensions, evaluation criteria, risk life cycle steps and governance levels — for deciding how much oversight an agent needs, and ties those decisions to the agent's autonomy, authority and context. It builds explicitly on OECD, NIST and ISO/IEC material, including ISO/IEC 23894 and ISO/IEC 42001, so assurance teams can map its pillars onto frameworks they already use, and it flags the security and accountability questions raised by external tool invocation.
❓ What’s Missing
The paper is conceptual: it offers dimensions, principles and example tables rather than templates, thresholds or sector-specific requirements, and it names no regulator or jurisdiction-bound obligation beyond citing GDPR and CCPA as compliance examples. The authors note that agent evaluation remains nascent, that emerging benchmarks are built for academic settings with predefined tasks and static environments, and that defining risk criteria and tolerability thresholds consistently remains a central challenge. Provider-side documentation requirements and quantitative acceptance criteria are described in general terms, and multi-agent ecosystems are presented as early-stage, so guidance for large-scale inter-organisational deployments stays at the level of failure modes.
👥 Best For
Adoption leads and decision-makers planning agent deployments, technical leaders designing evaluation and monitoring pipelines, and governance, risk and compliance teams translating classification and evaluation evidence into proportionate safeguards. Provider product and safety teams will find the evaluation principles and baseline governance mechanisms useful when documenting agent capabilities for enterprise customers.
📄 Source Details
AI Agents in Action: Foundations for Evaluation and Governance, a white paper published by the World Economic Forum in collaboration with Capgemini, dated November 2025. Contributors are listed in an acknowledgements section; no individual author byline is printed. It runs to 34 pages in English, with no series or reference number. No URL for the document itself appears in the text; endnote URLs point to cited external works. The full 34-page text extraction was available; figures were extracted as labels only.
