AI Agents

What are AI agents

AI agents are often referred to as agentic AI systems. An AI agent is a task-oriented system that follows predefined rules or prompts to carry out specific actions. In contrast, agentic AI denotes a more advanced, goal-driven architecture in which the system can autonomously plan, reason, and adapt to achieve high-level objectives with minimal human intervention. AI agents are a fast-growing class of AI-enabled systems that can act on behalf of users or organisations by perceiving a task environment, reasoning about what needs to be done, and executing multiple steps toward a goal. Unlike a conventional AI model that produces a single answer to a prompt, an AI agent can plan, use tools, interact with software, maintain context across steps, and complete workflows with limited human intervention. This shift from passive generation to active execution is what makes AI agents especially important for governance, regulation, and legal analysis.

The term “agentic AI” is increasingly used in international policy and technical literature to describe systems that exhibit goal-directed behaviour, task decomposition, tool use, interaction with humans or other agents, and persistence over time. The OECD’s conceptual work on agentic AI treats these features as the core of the landscape, while Singapore’s governance framework describes agentic systems as AI with autonomy and action-taking capabilities that can operate across multiple steps and invoke tools with limited human intervention. These definitions matter because they show that AI agents are not just chat interfaces or productivity aids; they are systems capable of acting in ways that can affect rights, operations, finances, and institutional decision-making.

AI agents are already being deployed across a wide range of settings. In enterprise contexts, they assist with document drafting, scheduling, data retrieval, software development, workflow automation, customer support, and compliance tasks. In more advanced settings, they may coordinate with other agents, manage sub-tasks over longer horizons, and use external services or APIs to complete assigned objectives. The AI Agent Index documents thirty prominent systems and shows that modern agents are already being built with varying levels of autonomy, control, ecosystem integration, and safety features. This growing deployment is significant because the more an AI system can act, the more important it becomes to know who is responsible for its behaviour, what permissions it has, how it is monitored, and how failures are detected and corrected.

From a legal and policy perspective, AI agents create a different set of questions from those raised by ordinary AI systems. A model that only predicts or generates text may still raise concerns about bias, privacy, or transparency, but an agent adds the possibility of direct action, including making changes to data, interacting with external systems, or triggering real-world consequences. That makes issues such as accountability, authorisation, logging, human oversight, data protection, and incident response much more urgent. Comparative governance scholarship also suggests that many existing AI frameworks remain model-centric and do not fully capture runtime behaviour, tool orchestration, or multi-agent interaction, leaving a gap between technical capability and legal control.

For this reason, AI agents are becoming a central topic in both international AI governance and enterprise risk management. The OECD AI Principles continue to provide the baseline for trustworthy AI, especially with their emphasis on human-centred values, transparency, robustness, and accountability. Singapore’s Model AI Governance Framework for Agentic AI offers one of the most detailed practical blueprints for organisations deploying such systems, while NIST’s AI risk management guidance and security-related profiles show how agentic systems can be folded into broader governance and incident-response practices. Together, these sources show that AI agents are no longer a future concept; they are becoming a present governance challenge that requires careful attention to definition, oversight, and institutional design.

Significance and common usage

In policy and research discussions, “agentic AI” is used to capture systems that go beyond single‑turn outputs to exhibit goal‑directed behaviour, task decomposition, tool use, interaction with humans or other agents, and persistence over time. OECD’s work on agentic AI emphasises these features and treats AI agents as a distinct category of AI systems because once AI can act, not just predict or generate, it raises qualitatively different questions about control, accountability, safety and data governance. As agentic AI begins to appear in enterprise workflows, technical operations and digital administration, “AI agents” are increasingly understood as operational actors within organisations rather than mere analytical tools, which explains why they are becoming a central focus of AI governance debates.

Official definition of AI agents

Indian legal and policy context

As of now, Indian legislation and case law do not provide a formal statutory or judicial definition of “AI agents” or “agentic AI,” and most domestic references to artificial intelligence continue to use broader terms such as “AI systems,” “AI technologies,” or “algorithms.” This is evident across Indian policy and regulatory documents, which focus on data protection, fairness, transparency, and automation in domains like finance, telecom, and e‑governance, but do not yet carve out a separate legal category for action‑taking, tool‑using AI systems. In practice, autonomy and automated decision‑making are treated as properties of AI systems generally, rather than as defining features of “agents,” and agentic behaviour—such as multi‑step tool use or workflow execution—is not explicitly recognised in statutory language or judicial doctrine. This means that Indian debates about AI agents currently rely largely on international conceptual work and industry practice rather than on explicit domestic legal definitions, creating a gap between emerging technical realities and formal legal terminology.

OECD conceptual foundations

Internationally, the most influential conceptual work on AI agents comes from the OECD’s report “The Agentic AI Landscape and Its Conceptual Foundations,” which reviews definitions of AI agents and agentic AI across technical, industry, and policy sources. The report identifies recurring features such as goal‑directed operation, task decomposition and planning, tool use, interaction with humans or other agents, and persistence over time, and then maps these features to the key elements of the OECD definition of an AI system. In this framework, AI agents are described as AI‑enabled systems that perceive inputs or an environment, reason about goals, plan and sequence actions, use external tools or services, and act with a degree of autonomy to achieve specific objectives for users or organisations. Agentic AI refers more broadly to AI systems that exhibit agentic characteristics—such as goal‑directed operation, task decomposition and planning, tool use, persistence over time, and a degree of autonomous action—whether implemented as a single agent or as multiple coordinated agents. Multi‑agent systems, in which several agents break down tasks, collaborate, and jointly pursue complex goals over time, should be understood as one prominent variation of agentic AI rather than as its defining feature. This framing recognises that even a single AI agent can display agentic behaviour through planning, tool use, and sustained autonomous action, while multi‑agent architectures represent a scalable extension for more complex, distributed workflows..

The OECD report does not impose a single legal definition, but it provides conceptual clarity for policymakers by distinguishing agentic systems from non‑agentic AI and highlighting governance‑relevant features such as autonomy, interaction, persistence, and tool orchestration. It situates AI agents within the OECD AI Principles, emphasising that even as systems gain more autonomy, human agency and oversight, technical robustness and safety, transparency and explainability, and accountability must remain central pillars of governance. In practice, this conceptualisation offers a reference point for any jurisdiction—including India—that wishes to define or interpret “AI agents” in a way that aligns with emerging international standards while retaining flexibility in domestic law.

Singapore’s operational definition

Singapore’s Model AI Governance Framework for Agentic AI (Version 1.0) provides a more operational, governance‑oriented definition of AI agents. The framework characterises agentic AI as “AI agents with autonomy and action‑taking capabilities” that can perform tasks across multiple steps, invoke tools, and interact with users and other systems with limited human intervention. It notes that such systems are typically built on large language models or other foundation models, but add orchestration layers—planning modules, tool‑calling components, memory, protocols, and sometimes multi‑agent configurations—that enable them to behave as semi‑autonomous digital workers within organisational workflows.

This definition is embedded in a four‑pillar governance blueprint that covers assessing and bounding risks upfront, making humans meaningfully accountable, implementing technical controls and processes across the lifecycle, and enabling end‑user responsibility through transparency and training. In effect, the framework defines AI agents not only by what they are, but also by what they require: limits on autonomy, least‑privilege access to data and tools, documented workflows and protocols, strong logging and monitoring, and clear escalation and approval paths for high‑impact actions. Because the framework is issued by Singapore’s Infocomm Media Development Authority (IMDA) as official guidance and was updated on 20 May 2026 (Version 1.5) to address multi‑agent systems and third‑party agents, it now functions as an official, non‑binding governance framework that serves as a key definitional reference for agentic AI in Singapore and is influencing how legal and policy actors in the region talk about AI agents. The Model AI Governance Framework for Agentic AI (MGF) is expressly advisory in nature—a “living document” that provides structured best practices for managing agentic AI risks but does not create binding statutory obligations.

NIST and delegated authority

In the United States, while there is no statutory definition of “AI agent,” the general NIST AI Risk Management Framework (AI RMF 1.0) treats AI systems—including agents—as tools that operate under delegated authority to perform tasks on behalf of humans, focusing on risk mapping, measurement, and management across the AI lifecycle. This broad framework does not prescribe specific technical controls for agent identity or authorisation but instead offers a flexible, voluntary risk-governance structure applicable to all AI systems.

Separately, NIST’s more recent agent-specific work—notably the February 2026 concept paper Accelerating the Adoption of Software and Artificial Intelligence Agent Identity and Authorization and the associated AI Agent Standards Initiative—proposes that, for governance purposes, AI agents be treated as identifiable entities within enterprise identity systems, with their own credentials and permissions enabling them to act within organisational environments. This initiative, led by NIST’s National Cybersecurity Center of Excellence (NCCoE), explores how existing identity and access management standards (such as OAuth, OpenID Connect, and attribute-based access control) can be adapted to provide secure identity, authorization, auditing, and non-repudiation for software and AI agents.

Commentary drawing on these NIST materials (and related ISO governance standards) recommends that organisations formally onboard, register, periodically review, and decommission AI agents, and that their actions be traceable and auditable with clear links to responsible owners and risk controls. Some commentators further suggest that, as a risk-mitigation practice, delegated authority for agents should generally be narrower than the authority of the human they support, though this is presented as a recommended control rather than a binding NIST, ISO, or legal requirement.

This functional approach does not create a legal definition in the sense of a statute, but it operationalises the concept of an AI agent for risk management and compliance: an AI agent is defined by its ability to act under delegated authority in ways that require identity and access management, monitoring, lifecycle governance, and accountability. For Indian regulators and courts, this perspective is particularly relevant because it aligns with existing legal notions of delegated decision‑making and organisational responsibility, and it suggests how AI agents can be conceptualised as actors whose permissions, boundaries and liabilities must be expressly defined and supervised within governance frameworks.

• Digital Personal Data Protection Act, 2023 (DPDP Act) The DPDP Act imposes accountability on Data Fiduciaries—entities that determine the purpose and means of processing personal data—for all processing activities, including those carried out by automated or AI systems. While the Act does not use the term “AI agent,” it requires that individuals not be subjected to decisions based solely on automated processing without meaningful human oversight, thereby mandating explainability, audit trails, and human‑in‑the‑loop controls for high‑stakes AI decisions. This statutory duty mirrors the governance proposal that AI agents’ actions must be traceable, auditable, and linked to responsible owners.
• Information Technology Act, 2000 (and proposed Digital India Act) Section 79 of the IT Act provides a safe‑harbour regime for intermediaries subject to due‑diligence obligations, which has been interpreted in cases like Shreya Singhal v. Union of India (2015) to require active oversight of platform‑hosted content. Commentators note that analogous principles could extend to AI systems acting on behalf of organisations, requiring clear delegation, monitoring, and liability attribution. The forthcoming Digital India Act is expected to introduce risk‑based classifications and enhanced accountability for AI and algorithmic systems, further reinforcing the need for defined permissions and boundaries.

Types of AI agents

Major categories of AI agents

The technical and policy literature identifies several broad types of AI agents, reflecting different roles, capabilities and risk profiles. OECD’s agentic AI landscape report maps systems such as personal and enterprise task assistants that manage administrative workflows, developer and operations agents that support code development and infrastructure management, business process agents embedded in compliance, KYC and back‑office tasks, research and analysis agents that autonomously search and synthesise information, and multi‑agent systems where multiple specialised agents collaborate or negotiate to achieve complex goals. The 2025 AI Agent Index complements this mapping by profiling thirty agent systems across categories, documenting their origins, architectures, capabilities, autonomy and control features, ecosystem roles and safety mechanisms, and thereby demonstrating that diverse agent types are already deployed in professional and personal domains.

Industry sources similarly distinguish productivity‑oriented assistants, workflow agents, coding and DevOps agents, customer‑facing service agents and complex multi‑agent ecosystems. IBM, AWS, SAP and BCG, for example, describe agents that can orchestrate multistep workflows, work together to solve complex business problems, and function as central coordination points for AI‑driven tasks. These categories are useful starting points for legal experts considering how different kinds of agents might intersect with sector‑specific regulation and risk.

Appearance in official databases

AI agents in Indian judicial and regulatory data

At present, AI agents do not appear as a distinct, systematically tagged category in Indian judicial or regulatory databases. Available case‑law repositories, regulatory portals and government reports tend to refer generically to “artificial intelligence,” “AI systems,” or “algorithms,” and focus on issues such as data protection, discrimination, algorithmic transparency, or automation in procedure rather than explicitly identifying agentic behaviour. Where AI is recorded in sectoral contexts—such as telecom or financial regulation—the emphasis is usually on the use of AI technologies in network optimisation, fraud detection or customer service, without separate tagging for agents that may perform multistep actions. This means that, as of now, official Indian databases do not systematically distinguish static AI systems from action‑taking agents, and any references to agents are likely to be buried within broader AI categories.

However, this observation is based on a review of publicly available secondary sources and expert commentaries rather than a comprehensive, documented search across all Indian judicial and regulatory databases. No single, publicly accessible metadata schema or taxonomy in Indian case‑law or regulatory repositories currently classifies decisions or guidelines by “AI agent” or “agentic AI” as a discrete category. A definitive, database‑wide claim would require a systematic search methodology—spanning platforms such as the Supreme Court’s e‑Courts repository, High Court databases, Manupatra, SCC Online, and regulatory portals of MeitY, RBI, TRAI, and SEBI—using controlled queries for terms such as “AI agent,” “autonomous system,” “agentic,” and related keywords, together with an audit of tagging and classification practices. Pending such a methodologically documented review, the statement should be understood as a qualified observation drawn from available literature and regulatory analyses rather than an exhaustive empirical finding.

International databases and inventories

Internationally, specialised technical and governance‑oriented databases have begun to capture AI agents more explicitly. The 2025 AI Agent Index, hosted by MIT and collaborators, functions as an early academic research repository documenting thirty agent systems across technical and safety dimensions, including origin, design choices, capabilities, autonomy and control, ecosystem integration and safety measures. Although it is not a judicial or governmental database, it is designed to support policymakers and researchers by providing structured information about agentic systems and their governance‑relevant characteristics. OECD’s AI risks and incidents work, while primarily focused on incidents and hazards, describes harms that can involve tool‑using and action‑taking systems and thereby indirectly touch agentic behaviour when cataloguing failures.Although it is not a judicial database, it is designed to support policymakers and researchers by providing structured information about agentic systems and their governance‑relevant characteristics. OECD’s AI risks and incidents work, while primarily focused on incidents and hazards, describes harms that can involve tool‑using and action‑taking systems and thereby indirectly touch agentic behaviour when cataloguing failures.

Research that engages with AI agents

Conceptual and technical research

OECD’s “The Agentic AI Landscape and Its Conceptual Foundations” is a key conceptual contribution, systematically analysing how AI agents and agentic AI are defined in technical, industry and policy sources, and mapping these features to the OECD definition of AI systems and AI Principles. It identifies goal‑directed behaviour, planning, tool use, interaction and persistence as central elements, and explores how these characteristics create new governance touchpoints for policymakers, including issues of autonomy, control, multi‑agent interaction and open‑ended operation. The report helps to establish a common vocabulary for discussing agentic systems and guides regulators in thinking beyond static model governance.

On the technical side, the AI Agent Index provides empirical data on thirty agentic systems, documenting their technical architectures, capability profiles and safety controls, including mechanisms for human oversight, logging, and constraint enforcement. This work shows how agentic AI systems are being constructed in practice and identifies trends in design, such as the use of planners, tool‑calling layers, memory components and multi‑agent configurations.

Governance‑focused research

Governance‑focused research, such as “Mind the Gap: How the Technical Mechanisms of Agentic AI Interact with Governance Frameworks”, examines how agent capabilities interact with existing AI laws, intergovernmental guidance and national regulatory frameworks across multiple continents. It finds that many current frameworks remain model‑centric, regulating training, data and static outputs, and do not fully capture runtime behaviour, action‑taking, tool orchestration or multi‑agent dynamics. The paper argues that agentic AI requires governance layers that focus on operational mechanisms—such as delegated authority, interaction with external systems and human‑in‑the‑loop structures—and suggests that regulators need to adapt incident reporting, accountability rules and risk‑tiering to recognise agents explicitly.

Singapore’s Model AI Governance Framework for Agentic AI translates these insights into practice, offering detailed lifecycle guidance on assessing and bounding risks, ensuring meaningful human accountability, implementing technical controls and processes, and enabling end‑user responsibility. Commentary on the framework from law and policy firms highlights its treatment of multi‑agent systems, third‑party agent risks, updated safety components and differentiations of roles across the agentic AI value chain.

Within the Indian justice context, direct research on AI agents in courts, tribunals or legal processes is still nascent, and existing work tends to focus on algorithms, automated decision‑making or evidence handling more broadly. However, the international research above provides conceptual and methodological tools that Indian scholars can use to explore how agentic systems might affect procedural fairness, access to justice, judicial administration and regulatory enforcement, and how they can be documented and analysed beyond official records.

International experiences

Singapore

Singapore offers one of the clearest examples of early agentic AI governance. The Model AI Governance Framework for Agentic AI (Version 1.0) provides a practical blueprint for organisations deploying AI agents with autonomy and action‑taking capabilities, structured around four core dimensions: assessing and bounding risks, ensuring meaningful human accountability, implementing technical controls and processes, and enabling end‑user responsibility. Updates issued in May 2026 refine components of agentic AI by adding safety and reliability elements such as access controls, guardrails, human approvals, logging and monitoring; expand discussion of risks arising from multi‑agent systems and third‑party agent usage; differentiate responsibilities between platform providers and system/app developers; and provide more detailed guidance on technical controls and change management. These experiences show how a regulator can quickly move from general AI governance to agent‑specific practical guidance.

OECD countries and intergovernmental guidance

In OECD countries, agentic AI is being integrated into AI policy through conceptual and competition‑focused work. The OECD agentic AI report provides a common conceptual foundation and encourages governments to consider agent autonomy, coordination and tool use when designing AI regulation. Other OECD papers examine how generative and agentic systems reshape competition in downstream markets, noting that agentic AI can lower barriers to entry and enable new forms of service delivery while also creating risks related to data access, vertical integration and algorithmic conduct. These experiences suggest that agentic AI is relevant not only to safety and rights, but also to economic regulation.

United States and enterprise governance

In the United States, NIST’s AI Risk Management Framework and Cybersecurity Framework Profiles for AI treat systems that perform tasks under delegated authority as part of organisations’ AI risk posture, recommending that incident‑response plans, logging practices and risk assessments be adapted to account for AI systems that act, not just predict. Enterprise‑oriented governance frameworks and security guidance discuss “agent sprawl,” the need for central agent inventories, observability, and oversight mechanisms as practical challenges for organisations deploying many agents in different workflows. These experiences show that agent governance is increasingly viewed as an operational discipline within firms, not just a matter for regulators.

Data challenges

Traceability and logging

AI agents create significant data challenges because their behaviour unfolds over sequences of actions rather than single outputs. To analyse or audit an agent, organisations need logs that capture inputs, plans, tool calls, state changes, outputs and human interventions across different systems and over time. Without such structured logging, reconstructing what an agent did, why it did so, and how an incident occurred can be extremely difficult, especially in multi‑agent environments. Existing data infrastructures in justice and regulation, which are often designed around static events and documents, may not yet be equipped to capture these dynamic chains of action.

Classification and schema limitations

Many current AI incident and system databases, including those used for regulatory reporting, were built for models and may lack fields for autonomy level, tool scope, multi‑agent interactions or runtime decision chains. Research on standardised schemas and taxonomies for complex AI systems suggests that new data structures are needed to represent agent characteristics and behaviours in a way that supports comparative and statistical analysis. Without such schemas, data about agents is likely to remain fragmented across proprietary logs, dashboards and project documentation, preventing systemic analysis of agent deployment and associated risks.

Privacy and visibility

Agents often access and transform personal and confidential data across tasks, so logs of their behaviour may themselves contain sensitive information that cannot easily be shared or aggregated without privacy concerns. At the same time, enterprises may deploy multiple agents without a central inventory or governance structure, leading to “agent sprawl” in which organisations lack a complete picture of where agents are active and what data they touch. Governance commentaries identify this lack of visibility as a major barrier to effective risk assessment and compliance because untracked agents can act in ways that are invisible to those responsible for oversight.

Way ahead

Standardising and harmonising agent data

To address these challenges, international governance work and enterprise practice recommend establishing central agent inventories and standardised logging frameworks. Agent inventories should record, for every deployed agent, key attributes such as purpose, autonomy level, tool access, data scope, deployment environment and responsible owners, thereby providing harmonised metadata that support both internal oversight and external reporting. Standardised logs should capture agent behaviour in a structured form, including inputs, plans, actions, outputs and interventions, so that data can be analysed across agents and time. Research on schemas and taxonomies, such as work on standardised AI system databases and the AI Agent Index, can inform the design of these data structures.

Improving data collection and lifecycle governance

Lifecycle governance frameworks, particularly Singapore’s MGF for Agentic AI, emphasise embedding data collection, logging and monitoring throughout design, development, deployment and post‑deployment rather than treating them as afterthoughts. This means planning from the start how agent behaviour will be recorded, how risks will be assessed, and how incidents will be detected and handled. Comparative research like “Mind the Gap” argues that governance needs to focus on runtime mechanisms and delegated authority, implying that data about agents should reflect their operational role and interactions, not only their training data or static capabilities. For India, adopting similar lifecycle and data‑centric approaches can help standardise and harmonise agent data, improve future data collection and facilitate systemic analysis.

Enabling systemic analysis

Senior judges, regulators, academics and research organisations internationally have begun to suggest that systemic analysis of AI agents will depend on integrating agent data into existing justice and regulatory data infrastructures. This includes linking agent inventories and logs to incident reporting systems, sectoral regulators’ databases and, where relevant, court and tribunal information systems. Doing so would enable oversight bodies to see patterns in agent deployment, correlate agent behaviour with incidents or complaints, and assess the impact of agentic systems on rights, fairness and institutional performance. In India, similar suggestions could be adapted to ensure that AI agents used in justice‑adjacent contexts are visible in data and amenable to analysis, thereby supporting more informed policy and judicial responses.

Senior judges, regulators, academics and research organisations internationally have begun to suggest that systemic analysis of AI agents will depend on integrating agent data into existing justice and regulatory data infrastructures.

• Judicial voices: Senior Judge Herbert Dixon Jr. (Superior Court of the District of Columbia) has participated in a 2025 national task force—launched by the nonpartisan Council on Criminal Justice (CCJ) in partnership with RAND—that is developing standards and evidence‑based recommendations for the integration and oversight of AI in the criminal justice system, including principles for linking AI inventories and incident reporting to oversight structures. U.S. judges participating in the 2026 AI Policy Consortium/NCSC–Thomson Reuters Institute webinar (e.g., Judge Schlegel) have similarly emphasised structured guardrails, verification protocols, and audit trails for judicial AI use, underscoring the need for traceable logs that can be reviewed by oversight bodies.
• Academics and research organisations: Researchers associated with the Partnership on AI and collaborating institutions have proposed mandatory/voluntary AI incident reporting regimes that incorporate standardised incident components—including technical data, affected entities, and contextual information—so that regulators and courts can correlate AI behaviour with harms and complaints. Parallel work on “Incident Analysis for AI Agents” (arXiv, 2025) proposes an incident analysis framework tailored to agents, explicitly to inform emerging incident reporting processes and enable systemic oversight of agentic systems.

Also known as

Comparable and synonymous terms

In Indian and international discourse, AI agents are commonly referred to as “AI agents,” “agentic AI systems,” “agentic AI,” “autonomous AI agents,” or “AI workflow agents.” Technical literature also uses terms such as “software agents,” “intelligent agents”, and “multi‑agent systems” to describe related concepts, though these predate current large‑model‑based agent architectures and may not fully capture modern capabilities such as LLM‑driven planning, tool use and persistent memory. Enterprise governance articles and frameworks increasingly speak of “delegated AI agents,” “enterprise AI agents”, and “agentic AI workflows” to highlight that these systems perform real work under organisational authority and therefore require specific oversight. As Indian law and policy evolve, clarifying which of these terms are used in official documents, and how they relate to one another, will be important to avoid confusion and ensure consistent treatment of agentic systems.

References

Contributors & review status

Curated / edited bynot yet recorded

Peer reviewpending

Project reviewpending

Expert reviewpending

Page created 2026-08-21  ·  Last updated 2026-08-23