Cross-corpus research synthesis
Agentic AI Governance
Obligations specific to AI systems that take autonomous multi-step actions (browse, transact, plan, recurse). Distinct from foundation_models (capability) and catastrophic_risk (outcome) — this is the action-surface frame. Surfaces in EU AI Office GPAI Code drafts, UK AISI agent evaluations, Seoul Frontier AI Safety Commitments §3, NIST AI 600-1.
Synthesised deterministically from 20 articles that engage this theme. Empirical consensus: emerging · contested: Should governance attach to the AGENT (multi-step actions, tool use, recursion) or to the model that powers it? Capability-tier vs action-tier frames are unresolved across jurisdictions.. Full theme article: /wiki/agentic-systems-governance. Machine-readable: /wiki/synthesis.json.
Cross-jurisdiction stances (5 govern, 18 engage)
| Instrument | Verdict | Provision excerpt / citation |
|---|---|---|
| EU AI Act | implicit | “Deployers of high-risk AI systems shall take appropriate technical and organisational measures to ensure they use such systems in accordance with the instructions for use accompanying the systems…” Arts. 26-29 deployer obligations apply to agent operators; Arts. 51-55 GPAI obligations capture the underlying model |
| Interim Measures for Generative AI Service Management | implicit | Arts. 4, 8 (service-provision scope) — agent-like generative services fall within registration + safety-assessment obligations |
| G7 Hiroshima AI Process Code of Conduct | implicit | Code §1 'advanced AI systems' + §3 risk-identification cover agentic behaviour through capability frame |
| Council of Europe Framework Convention on AI | implicit | General-AI scope (Art. 3) covers agent systems; no agent-specific provision |
| NIST AI Risk Management Framework | implicit | “Mechanisms are in place and applied, and responsibilities are assigned and understood, to supersede, disengage, or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use.” Map / Manage functions apply to autonomous systems; no agent-specific profile yet |
| Bletchley Declaration on AI Safety | implicit | Frontier-AI risk frame includes autonomous-action risks; no specific obligation |
| Seoul Declaration on Safe, Innovative and Inclusive AI | governs | Frontier AI Safety Commitments §3 — pre-deployment capability evaluations include agentic behaviours under 'realistic deployment conditions' |
| NIST AI RMF Generative AI Profile | governs | NIST AI 600-1 names Value Chain + Component Integration as risk category covering agentic / tool-use deployments |
| Brazil AI Bill (PL 2338/2023) | implicit | Risk-based framework (PL 2338 Arts. 13-15) covers agent systems under high-risk tiers if applicable |
| Anthropic Responsible Scaling Policy (RSP) v2 | governs | RSP v2 — ASL thresholds include 'autonomous AI replication' + agentic capability evaluations |
| OpenAI Preparedness Framework | governs | Preparedness Framework — Model Autonomy is one of four named risk categories |
| Google DeepMind Frontier Safety Framework | governs | FSF Critical Capability Levels — Autonomy is one of four named CCL domains |
| Meta Frontier AI Framework | implicit | Capability tiers cover agentic behaviour; not named as a distinct category |
| UK-US AI Safety Institute Memorandum of Understanding | implicit | Joint AISI capability evaluations include agentic-behaviour testing |
| California SB-53: Transparency in Frontier Artificial Intelligence Act (TFAIA) | implicit | Bus. & Prof. Code § 22757.11 catastrophic-risk prongs cover a model acting 'without meaningful human oversight' or 'evading the control of its developer or user' (§ 22757.13 incident reporting); reached only via the catastrophic-risk lens, not a dedicated agentic-autonomy regime |
| Revised Product Liability Directive (Directive (EU) 2024/2853) | implicit | Art. 7(2)(c) — defectiveness accounts for a product's ability to continue to learn or acquire new features after market placement; Art. 11(2) — post-placement software-update liability within the manufacturer's control |
| Directive (EU) 2024/2831 on improving working conditions in platform work | implicit | Automated decision-making systems that autonomously allocate tasks, set pay, monitor and discipline platform workers function as agentic management tools; the Directive subjects them to operative tran (paraphrase) Directive (EU) 2024/2831, Articles 9-11 |
| New York RAISE Act: Responsible AI Safety and Education Act | implicit | N.Y. Gen. Bus. Law § 1420(7) critical harm includes model conduct 'with no meaningful human intervention'; § 1420(13) 'safety incident' includes autonomous model behaviour + control failures — autonomy reached via the catastrophic-risk/incident lens, not a dedicated agentic regime |
Evidence convergence
Sources the corpus cites for this theme across multiple articles — a scientometric consensus signal computed from inline prose citations (the more articles independently cite a source, the more load-bearing it is for this theme). 29 sources are cited by ≥2 articles.
- 14×Governing AI Agents — cited by 14 articles
- 13×An interdisciplinary account of the terminological choices by EU policymakers ahead of the final agreement on the AI Act: AI system, general purpose AI system, foundation model, and generative AI — cited by 13 articles
- 13×Multi-Agent Risks from Advanced AI — cited by 13 articles
- 12×The EU model of AI governance: regulating artificial intelligence through law and policy — cited by 12 articles
- 11×Infrastructure for AI Agents — cited by 11 articles
- 11×Artificial intelligence and synthetic biology: biosecurity risks, dual-use concerns, and governance pathways — cited by 11 articles
- 8×Two types of AI existential risk: decisive and accumulative — cited by 8 articles
- 7×Missing the Mark: Adoption of Watermarking for Generative AI Systems in Practice and Implications Under the New EU AI Act — cited by 7 articles
- 6×Generative AI and data protection — cited by 6 articles
- 6×AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents — cited by 6 articles
- 6×Defending Compute Thresholds Against Legal Loopholes — cited by 6 articles
- 6×International Agreements on AI Safety: Review and Recommendations for a Conditional AI Safety Treaty — cited by 6 articles
- 4×Audio deepfakes and the regulation of the landlords of creativity — cited by 4 articles
- 4×Identifying Algorithmic Decision Subjects' Needs for Meaningful Contestability — cited by 4 articles
- 3×Open Foundation Models and TDM Exceptions to Copyright – Building Blocks for an AI Ecosystem — cited by 3 articles
- 3×Artificial Intelligence and Nuclear Weapons Proliferation: The Technological Arms Race for (In)visibility — cited by 3 articles
- 3×A Teleological Interpretation of the Definition of DeepFakes in the EU Artificial Intelligence Act—A Purpose-Based Approach to Potential Problems With the Word 'Existing' — cited by 3 articles
- 3×Authenticated Delegation and Authorized AI Agents — cited by 3 articles
- 2×arxiv:2504.18236 — cited by 2 articles
- 2×Visibility into AI Agents — cited by 2 articles