Loading
An AI agent is a software system that uses an AI model, typically a large language model, to pursue a goal with a degree of autonomy: it plans a sequence of steps, decides which tools or data sources to call, observes the results, and adjusts its next action, rather than executing a single fixed script written in advance by a developer. This is the property the U.S. National Institute of Standards and Technology (NIST) captures in its technology-neutral definition of an "AI system" as one designed to operate "with varying levels of autonomy," and it is why NIST opened a dedicated AI Agent Standards Initiative in February 2026 rather than treating agents as ordinary software. For governance and compliance teams, that autonomy is the whole story: an agent that can independently send a message, execute a purchase, modify a record, or call another system is capable of taking actions with real-world consequences that a chatbot or a traditional rules-based automation script cannot take on its own. Because agentic systems moved into mainstream enterprise tools through 2025, six national cyber-security agencies, including CISA and the NSA, issued the first coordinated international security guidance for agentic AI in May 2026, treating it as a distinct risk category that needs its own controls rather than an extension of ordinary AI or RPA governance.
Run the free AI Health CheckAI Agent, a software system that uses an AI model, most often a large language model, to autonomously plan and carry out a sequence of actions toward a goal, invoking tools, retrieving data, and adapting its next step based on what it observes, rather than following one fixed, pre-programmed script.
The defining feature of an AI agent is not intelligence but autonomy: unlike a chatbot that answers once and stops, or robotic process automation that repeats the same scripted steps, an agent decides for itself which action to take next and can keep acting for extended periods without a person checking each step. That autonomy is exactly what has drawn regulatory attention: NIST opened a dedicated AI Agent Standards Initiative in February 2026, and in May 2026 CISA, the NSA and allied cyber agencies issued the first joint international security guidance built specifically for agentic AI, both treating autonomous action as a distinct governance risk requiring its own controls, especially mandatory human approval before irreversible or high-stakes actions.
Source: NIST AI Agent Standards Initiative (2026); Five Eyes joint guidance 'Careful Adoption of Agentic AI Services' (2026); NIST AI RMF 1.0 / NIST AI 100-1 (2023)
NIST AI Agent Standards Initiative (Feb 2026)
NIST's Center for AI Standards and Innovation (CAISI) launched this initiative on 17 February 2026 to develop industry-led technical standards and open protocols for agent security, identity and credentialing, and interoperability, treating agents as needing standards of their own rather than relying only on NIST's older general-purpose AI system definition.
Five Eyes joint security guidance (May 2026)
CISA, the NSA, Australia's ASD ACSC, Canada's Centre for Cyber Security, New Zealand's NCSC and the UK's NCSC jointly published 'Careful Adoption of Agentic AI Services' on 1 May 2026, the first coordinated multi-government guidance written specifically for agentic AI. It sets out five risk categories (privilege, design/configuration, behavioral, structural, and accountability) and states that human approval is a required control, not an optional one, before financial transactions, identity or access-control changes, external communications sent on a person's behalf, and any action that cannot be undone without significant effort.
ISO/IEC 22989:2022 terminology
The international AI terminology standard underlying much of this guidance defines an agent as an entity that senses and acts on its environment to achieve goals, and separately defines 'autonomy' (clause 3.1.5) narrowly as a system capable of modifying its own intended domain of use or goal without external intervention, control, or oversight, a much stricter bar than the colloquial 'autonomous AI agent' usage common in industry.
Industry technical framing
Anthropic's own engineering guidance distinguishes 'workflows', LLMs and tools orchestrated through predefined code paths, functionally close to RPA, from 'agents,' where the model itself dynamically directs which tools to call and in what order, echoing the same autonomy line regulators are now drawing.
Not all systems marketed as 'AI agents' carry the same governance risk, because autonomy is a spectrum rather than a single on/off property. NIST's baseline AI system definition already anticipated this, describing 'human-AI configurations spanning from fully autonomous to fully manual.' At one end sit systems where a human approves every single action; further along, a human monitors and can intervene but does not approve each step; at the far end, a system can operate for extended periods, hours or days, with no per-action check-in at all.
This distinction matters practically because the controls that make sense for a low-autonomy assistant are inadequate for a high-autonomy agent that has already acted before anyone reviews it. That is precisely the gap the 2026 Five Eyes guidance targets: it does not ask organizations to avoid agentic AI, but to match the strength of oversight, approval gates, scoped permissions, monitoring, and kill-switches, to how much autonomy a given deployment actually has, escalating controls as autonomy increases rather than applying one static policy to every system labelled an 'agent.'
What is the simplest definition of an AI agent?
An AI agent is a software system built around an AI model, typically a large language model, that plans and carries out a sequence of actions, calling tools, retrieving data, taking steps in other systems, to reach a goal, adjusting its approach as it goes rather than following one fixed, pre-written script.
How is an AI agent different from a chatbot?
A chatbot generates a single conversational response to a prompt and stops there; a person decides what happens next. An AI agent keeps going after that first response, it can decide on its own to call a tool, query a database, send a message, or take a further step toward a goal without a person prompting each individual action.
Is robotic process automation (RPA) the same thing as an AI agent?
No. RPA executes a fixed, pre-programmed script, the same steps in the same order every time, following explicit if-then rules. An AI agent uses an AI model to decide, at run time, which steps to take and in what order, which lets it handle variation an RPA bot would break on, but also makes its behavior harder to fully predict in advance.
Is 'agentic AI' the same as an 'AI agent'?
The terms are closely related but not always used identically. 'AI agent' usually refers to a specific instance of an autonomous system built for a task; 'agentic AI' is the broader umbrella term, used by NIST and industry alike, for the category of AI systems and architectures, including multi-agent systems, that exhibit that autonomous, tool-using behavior.
Is there official government guidance specifically for agentic AI?
Yes. NIST opened its AI Agent Standards Initiative in February 2026 to develop technical standards for agent security and interoperability, and in May 2026 CISA, the NSA, and cyber agencies from Australia, Canada, New Zealand and the UK jointly published 'Careful Adoption of Agentic AI Services,' the first coordinated international security guidance written specifically for agentic AI deployments.
Does an AI agent need human approval before it acts?
Regulators increasingly say yes for certain categories of action. The 2026 Five Eyes guidance states that human approval is a required control, not optional, before an agent executes a financial transaction, changes identity or access permissions, sends external communications on a person's behalf, or takes any action that cannot be easily undone.
Last reviewed July 2026
This page is general information about What Is an AI Agent?, not legal, regulatory, or professional advice, and does not capture every nuance or exception. Requirements change and can be fact-specific. Always verify against primary sources and your own qualified legal counsel before relying on it.