Governance practices, policy papers, and frameworks specific to AI agents and autonomous systems.
19 resources
Shavit et al. at OpenAI propose seven practices for governing agentic systems: evaluating suitability, constraining action space, setting default behaviours, ensuring legibility, automatic monitoring, attributability, and interruptibility. Foundational reference for later governance frameworks.
Chan et al. (Centre for the Governance of AI) argue agents need shared infrastructure - identity, credentials, reputation, action logs, dispute resolution - analogous to financial and web plumbing. Maps building blocks to governance goals.
Joe Kwon (Center for AI Policy) policy paper on governing autonomous AI, examining levers like capability evaluations, pre-deployment approvals, liability rules, and compute governance. Focused on US federal policy options for increasingly autonomous systems.
Szpruch et al. propose a scalable runtime-governance architecture for agents in financial services, with policy engines, audit trails, and real-time guardrails tied to regulatory obligations. Worked examples cover algorithmic trading and customer-facing agents.
Singapore IMDA's Model AI Governance Framework dedicated to agentic AI, translating its general MGF into agent-specific practices across five dimensions: oversight, traceability, reliability, interaction, and ecosystem. Aligned with the country's Companion Guide.
World Economic Forum white paper offering foundations for evaluating and governing enterprise agents, covering definitions, capability assessment, risk taxonomies, and oversight mechanisms. Synthesises practitioner input from WEF's AI Governance Alliance.
Institute for AI Policy and Strategy field guide cataloguing governance interventions across the agent lifecycle - alignment research, evaluations, deployment constraints, and incident response. Written for policymakers and deployers navigating overlapping proposals.
Kraprayoon et al. arXiv version of the IAPS field guide, systematising governance levers for AI agents across technical, organisational, and policy layers. Provides a taxonomy of interventions with maturity ratings and open research questions.
Kasirzadeh and Gabriel propose a four-dimensional characterisation of AI agents - autonomy, goal complexity, generality, and sociality - and use it to structure alignment and governance questions across today's and future systems.
OWASP v1.0 state-of-the-field report on agentic AI security and governance in 2025, summarising threat models, vendor controls, standards activity, and practitioner surveys. Identifies gaps between attacker capabilities and defender tooling.
Julia Smakman (Ada Lovelace) policy briefing assessing whether UK law currently regulates AI advisers, agents, and companions. Maps data protection, consumer, financial, and professional-services regimes against agent use cases and identifies gaps.
Ada Lovelace Institute press briefing summarising AWO's legal analysis of advanced AI assistants, which concludes existing UK consumer, equality, and data-protection laws do not adequately cover agent-specific harms and calls for targeted legislative reform.
Clifford Chance analysis of contractual and tort liability gaps created by agents acting autonomously across systems. Examines principal-agent doctrine, software supplier liability, and data-processing contracts, with drafting suggestions for enterprise agent deployments.
Tomasev et al. (Google DeepMind) formalise delegation as a governance mechanism for autonomous AI, modelling principal-agent dynamics and proposing protocol designs that keep humans in control as agents handle more decisions on their behalf.
Feng et al. propose a five-level scheme for classifying AI agent autonomy, adapted from SAE driving levels, tied to oversight requirements and evaluation tests at each level. Intended to standardise how deployers describe autonomy claims.
Noam Kolt's legal scholarship applies principal-agent theory from law and economics to AI agents, identifying information asymmetry, incentives, and loyalty problems. Proposes governance levers including disclosure duties, fiduciary analogues, and liability allocation.
Hadfield-Menell's Berkeley technical report formalises AI alignment as a principal-agent problem with incomplete contracts, drawing on mechanism design. Introduces inverse reward design and cooperative inverse reinforcement learning as alignment approaches.
Kandikatla and Radeljic propose a risk-based framework for calibrating human oversight of AI, tying oversight intensity to capability, context, and consequence. Maps oversight patterns (in-the-loop, on-the-loop, out-of-the-loop) to concrete risk tiers.
Harry Farmer (Ada Lovelace) report on policy dilemmas raised by advanced AI assistants: concentration of power, undermining of user agency, and regulatory fragmentation. Proposes UK policy responses across competition, consumer, and data protection.