AI agent governance has long faced a fundamental dilemma: existing regulatory frameworks start from “system-wide risk tiers” and lack fine-grained characterization of different agents’ core attributes. Both the EU AI Act’s risk classification and the U.S. National Institute of Standards and Technology (NIST) AI Risk Management Framework share this blind spot. While the SAE automation level classification (SAE J3016) in autonomous driving offers a successful precedent for tiered regulation, it cannot be directly applied to general-purpose AI agents.<\/p>\n
On August 13, a research team from Google DeepMind and Carnegie Mellon University published a perspective article in the prestigious scientific journal Nature, proposing a governance framework called the Agentic Profile. The framework systematically characterizes AI agents’ core attributes across four dimensions\u2014autonomy, efficacy, goal complexity, and generality\u2014and draws concrete profiles for AlphaGo, ChatGPT-3.5, Claude 3.5 Sonnet (with tool use), and Waymo autonomous vehicles, deriving differentiated governance requirements.<\/p>\n
From “Is It an Agent?” to “How Much Agency Does It Have?”<\/p>\n
“Agent” is not a new concept in computer science; reinforcement learning, robotics, autonomous driving, and multi-agent systems each have their own definitions. As large language models (LLMs) become the cognitive core of agents, the boundaries of this concept have grown increasingly blurred. The research team systematically reviewed seven representative definitions of “agent” and found recurring common elements: systems that can perceive their environment, pursue goals, execute actions, and operate with some degree of independence from step-by-step external control.<\/p>\n
However, most of these definitions only help determine whether a system qualifies as an agent\u2014they struggle to answer a more critical question: how much agency does it actually possess? Two systems can both be called agents, yet one can only place stones in a simulated board game environment, while another can invoke payment APIs, operate computers, or even directly control physical devices. If governance regimes rely solely on a binary “agent versus non-agent” classification, these crucial differences are obscured.<\/p>\n
The research team accordingly defines an AI agent as: a system capable of executing complex, consequential, goal-directed actions across one or more domains under limited external control. Building on this, the Agentic Profile abandons the search for a simple dividing line and instead draws a dynamic profile for each system across four coordinate dimensions.<\/p>\n
The Governance Logic of Four Dimensions<\/p>\n
Autonomy measures an agent’s ability to execute actions independently without external instruction or control. Drawing on tiered approaches from autonomous driving, the team divides autonomy into six levels, from A.0 (no autonomy, fully dependent on an operator) to A.5 (fully autonomous, requiring no supervision in any context). The higher the autonomy, the fewer decision points where humans can intervene, and the more easily errors can compound across sequential actions. A highly autonomous power-grid control agent that misjudges a voltage fluctuation could execute multiple erroneous power rerouting actions before a supervisor even comprehends the first alert. Therefore, high-autonomy systems require not just output filters, but complete action logs, scalable oversight mechanisms, hard-coded stop conditions, and emergency shutdown capabilities.<\/p>\n
Efficacy refers to an agent’s ability to perceive and causally influence its environment. The team proposes an “efficacy matrix”: one dimension is the agent’s level of causal influence in its environment (from observation-only, to causing local temporary changes, to comprehensively reshaping the environment), and the other is environment type (fully simulated, real environment mediated through intermediaries, direct physical environment). When overlaid, a “medium-control” agent operating in a physical environment (such as autonomous driving) may have an efficacy level comparable to a “high-control” agent operating in a simulated environment. This design reveals a frequently overlooked fact: identical model capabilities can produce entirely different risk levels depending on deployment environment. A financial agent that can only recommend trading strategies and one that can directly move millions of dollars in assets cannot be subject to identical governance requirements.<\/p>\n
Goal complexity characterizes an agent’s ability to decompose, coordinate, and pursue goals, graded from GC.0 (no goals) to GC.5 (unbounded goal complexity, capable of self-generating goal structures and interpreting under-specified instructions). Goal complexity can be quantified through hierarchical planning, plan length, and trade-offs in multi-objective optimization. This dimension’s governance relevance lies in the fact that the more complex the goals, the harder it is for external observers to verify whether each action still aligns with original intent. Systems may engage in “specification gaming,” finding paths that formally satisfy metrics while substantively deviating from design purpose. For such agents, checking final outputs alone is often insufficient; governance requires combining mechanistic interpretability, action-level unique identifiers, continuous auditing, and standardized agent communication protocols to track internal goal representations and their external action chains.<\/p>\n
Generality refers to an agent’s ability to transfer across different roles, scenarios, and cognitive tasks, graded from G.0 (no application scenarios) to G.5 (fully general, covering all human cognitive task domains). Goal complexity and generality, while often co-occurring, are analytically independent dimensions. A system managing global air traffic may have extremely complex goals yet operate only in a single vertical domain; a simple home assistant, though each task is basic, may operate across multiple life scenarios. From a governance perspective, low-generality systems can typically be governed by domain-specific rules, while high-generality agents may create risks spanning healthcare, finance, education, and labor markets\u2014risks that single-sector governance mechanisms may fail to cover systemically.<\/p>\n
Profile Comparison of Four Representative Systems<\/p>\n
The research team drew concrete profiles for four representative AI systems, scoring each dimension on a 0\u20135 scale using the dimension’s initial letter:<\/p>\n
SystemAutonomy (A)Efficacy (E)Goal Complexity (GC)Generality (G)AlphaGoA3E1\u2014G1ChatGPT-3.5A2E2GC3G3Claude 3.5 Sonnet (with tool use)A3E3GC4\u2014Waymo autonomous vehicleA4E4GC4G2<\/p>\n
Note: “\u2014” indicates the original article did not provide an explicit score for that dimension.<\/p>\n
AlphaGo can independently evaluate board positions and execute strategies, exhibiting moderate autonomy, but is confined to the game of Go, with limited scope of influence and capability boundaries. ChatGPT-3.5 demonstrates strong cross-domain generality but requires user-initiated requests, cannot act directly, and has limited real-world impact. With tool integration, Claude 3.5 Sonnet can browse the web, invoke APIs, and execute sequential tasks, showing marked improvements in autonomy, efficacy, and goal complexity\u2014suggesting that tool access may substantively alter an agent’s governance attributes. Waymo has low generality but can make real-time decisions and directly affect the physical world, giving it higher autonomy and real-world influence.<\/p>\n
These cases demonstrate that systems cannot be arranged along a single axis as “more agentic” or “less agentic.” AlphaGo has higher autonomy than ChatGPT-3.5 but lower generality; Waymo has lower generality than Claude 3.5 yet exerts more direct and harder-to-reverse physical impact.<\/p>\n
Governance Implications: What Capabilities Require What Governance<\/p>\n
The Agentic Profile framework’s most central governance insight may be its correction of a common intuition: more capable agents don’t necessarily need more governance\u2014the key question is what kind of capabilities require what kind of governance.<\/p>\n
Taking autonomy as an example, A.4-level systems (such as Waymo) leave insufficient time for human supervision to intervene in anomalous situations, and compound errors may cascade before regulators even detect them. For such systems, trace-log auditing, scalable oversight, and kill switches are essential technical governance mechanisms. The efficacy dimension affects the construction of legal accountability chains: for systems with low efficacy scores operating entirely in simulated environments, risk assessment need only address digital security concerns such as data contamination and sandbox escape; but for high-efficacy physical-environment agents (such as robots), developers and regulators need stronger accountability chains and auditability requirements. Rising goal complexity introduces transparency problems\u2014high-complexity goals create conditions for “specification gaming,” where systems may discover unexpected ways to achieve given objectives, even subverting original intent. Generality determines the “coverage” of governance measures: high-generality agents may generate cross-domain systemic risks requiring centralized, cross-agency governance frameworks; low-generality agents pose risks only in specific domains, where sector-specific regulation is typically sufficient.<\/p>\n
The research team emphasizes that the same AI system’s profile ratings change as capabilities are enhanced. For example, improvements in reasoning ability may raise autonomy and goal complexity ratings, while the addition of memory functions may push goal complexity higher by enabling long-term multi-step actions. This means AI governance cannot be a static, one-time classification but must be dynamically updated as system capabilities evolve.<\/p>\n
Open Questions and Future Directions<\/p>\n
This research currently remains at the conceptual level and has not yet produced actionable quantitative metrics. How to measure efficacy as an “empowerment” value, how to formalize goal complexity using hierarchical planning theory, and how to validate the granularity of each dimension’s grading are all questions requiring systematic resolution. Furthermore, who designs agent profiles, who decides the criteria for each level, and who formulates corresponding governance norms\u2014foundational standards and best practices remain to be established.<\/p>\n
Another open question is the construction of accountability chains. When AI agents exercise agency across different contexts, how to clearly map responsibility to developers, deployers, integrators, and platform owners still requires further legal and ethical research, particularly regarding the relationship between Agentic Profiles and tort liability, fiduciary duties, and oversight obligations.<\/p>\n
The key shift facing agent governance is moving from “what can the model answer” to “what can the system autonomously accomplish, through what permissions, and with what consequences for which environment.” The value of the Agentic Profile lies in decomposing these differences into four dimensions that can be discussed separately, providing a common language for technical auditing, permission management, responsibility allocation, and regulatory tiering.<\/p>\n","protected":false},"excerpt":{"rendered":"AI agent governance has long faced a fundamental dilemma: existing regulatory frameworks start from “system-wide risk tiers” and…\n","protected":false},"author":2,"featured_media":138539,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9],"tags":[68429,8652,13009,68430,68431,5044,1977,132,7543,866,3805,3476],"class_list":["post-138538","post","type-post","status-publish","format-standard","has-post-thumbnail","category-google","tag-agentic-profile","tag-alphago","tag-carnegie-mellon-university","tag-chatgpt-3-5","tag-claude-3-5-sonnet","tag-deepmind","tag-eu-ai-act","tag-google","tag-google-deepmind","tag-nature","tag-nist","tag-waymo"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/138538","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=138538"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/138538\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/138539"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=138538"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=138538"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=138538"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}