AI agent
A software system that pursues a goal by choosing its own steps and tools, usually built around a language model that calls tools in a loop.
- OpenAI’s Operator, which filled in and submitted orders in its own browser
- BabyAGI, which creates, ranks and completes its own tasks
- A coding tool that patches a reported bug and reruns the tests
- Earliest AI use
- 1950s
- Alan Kay credits John McCarthy with the idea and Oliver Selfridge the word.
- Current meaning
- 2023
- BabyAGI and AutoGPT run GPT-4 in loops toward goals that users set.
- Opposite
- Workflow
- Anthropic: model and tool calls follow paths fixed in code.
In plain words
AI agents are software systems in which a language model works toward a goal by choosing its own steps and tools, such as web search, a browser or code.
Given a goal, such as booking a table or fixing a software bug, an AI agent works out the steps itself. A large language model picks an action and uses a tool, such as a search engine or a code runner. It checks the result and decides what to do next. It stops when it judges the job done or reaches a limit set by its developers.1
OpenAI’s Operator, released in January 2025, was one of the company’s first agents. It typed, clicked and scrolled in its own web browser, and it handed logins and payments back to the user.2 In July 2025 OpenAI folded it into ChatGPT as ChatGPT agent.
The word predates language models. In Stuart Russell and Peter Norvig’s textbook, AI itself is the study of agents: systems that perceive their surroundings and act on them.3 In 2023, projects such as AutoGPT and BabyAGI applied the word to GPT-4 running in a loop, and that sense took over. The developer Simon Willison crowdsourced 211 definitions before concluding in 2025 that a shared meaning had emerged.4
An AI agent differs from a chatbot, which answers one message and waits for the next. It also differs from a fixed workflow, in which developers set every step in code. The adjective agentic describes how far any system acts on its own.
Tests in 2025 showed clear limits. In early 2025, METR found that leading models succeeded half the time at tasks that take a person about 50 minutes. That length had doubled about every seven months since 2019.5 In a Carnegie Mellon test, agents got office tasks wrong about 70% of the time. Gartner predicted in June 2025 that more than 40% of agentic AI projects would be canceled by the end of 2027.6
How it works
Most current agents run the same loop. A program sends the model a goal, instructions and short descriptions of the tools it may use. The model replies with either a final answer or a request to run a tool, a step called tool use. The program runs the tool, adds the result to the model’s context window and asks again. The ReAct method of 2022 has the model write out its reasoning between actions, so each step can respond to what the last one found. Tools can be connected through the Model Context Protocol.
Each result gives the model evidence of its progress, such as a test that now passes. Developers commonly cap the number of steps. Every step adds cost, and an early mistake can carry into every later step.1
An agent that reads web pages or email can also meet prompt injection: instructions planted in that content to redirect it. Operator was built to ignore such instructions. It also asked for approval before significant actions, such as sending an email, which keeps a human in the loop.2
Origin and first known use
From Latin agere, “to do, perform”. In English an agent has been one who acts since the late 1400s, and a deputy acting for another since the 1590s. By Alan Kay’s 1984 account, John McCarthy conceived of a computer agent in the mid-1950s, and Oliver Selfridge coined the word a few years later. They pictured a program that, given a goal, would carry out the steps and ask for advice when stuck. Software agents were a research field by the 1990s. From 2023 the phrase came to mean a language model that uses tools in a loop.
- 1590s Agent, used in English since the late 1400s for one who acts, takes on the sense of a deputy who acts for another.Online Etymology Dictionary, “agent” Etymonline →: source for 1590s
- 1950s In Alan Kay’s account, John McCarthy has the idea of a program that pursues a goal and asks for advice when stuck. Oliver Selfridge coins the term agent.Alan Kay (1984), quoted in Jeffrey M. Bradshaw, “An Introduction to Software Agents”, Software Agents, AAAI Press, 1997 umbc.edu →: source for 1950s
- 1995 Wooldridge and Jennings write that agent resists a single accepted definition, and propose a weak notion of agency: autonomy, social ability, reactivity and pro-activeness.Wooldridge and Jennings, “Intelligent Agents: Theory and Practice”, Knowledge Engineering Review 10(2), June 1995 ox.ac.uk →: source for 1995
- 2022 Yao and colleagues’ ReAct has a language model alternate reasoning with actions that gather information from outside sources, updating its plan as results arrive.Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, arXiv, 6 Oct 2022 arXiv 2210.03629 →: source for 2022
- 2023 Yohei Nakajima releases BabyAGI, in which GPT-4 creates, ranks and completes its own tasks toward an objective. With AutoGPT, it spreads the current sense.Yohei Nakajima, “Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications”, 28 Mar 2023 yoheinakajima.com →: source for 2023
- 2024 Anthropic separates workflows, where code fixes the order of model and tool calls, from agents, where the model directs its own process.Anthropic, “Building effective agents”, 19 Dec 2024 anthropic.com →: source for 2024
- 2025 OpenAI releases Operator, an agent that works in its own web browser. In July it is folded into ChatGPT as ChatGPT agent.OpenAI, “Introducing Operator”, 23 Jan 2025, with an update of 17 Jul 2025 openai.com →: source for 2025
- 2026 NIST publishes its AI Agent Standards Initiative, which works toward voluntary guidelines on the security, identity and interoperability of agents.NIST, “AI Agent Standards Initiative”, 17 Feb 2026 nist.gov →: source for 2026
In the wild
-
“…we propose a novel task-driven autonomous agent that leverages OpenAI’s GPT-4 language model, Pinecone vector search, and the LangChain framework…”
-
“As AI agents like Auto-GPT speed up generative AI race, we all need to buckle up | The AI Beat”
-
“In a LLM-powered autonomous agent system, LLM functions as the agent’s brain, complemented by several key components…”
-
“…large language models (LLMs) have demonstrated remarkable potential in achieving human-level intelligence. This has sparked an upsurge in studies investigating LLM-based autonomous agents.”
-
“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
-
“Operator is one of our first agents, which are AIs capable of doing work for you independently…”
OpenAI, product announcement openai.com →: OpenAI, product announcement, 23 Jan 2025Archived →: copy of OpenAI, product announcement, 23 Jan 2025 -
“AI agents get office tasks wrong around 70% of the time, and a lot of them aren’t AI at all”
-
“An LLM agent runs tools in a loop to achieve a goal.”
-
“…there’s some over-prediction going on in the industry. In my mind, this is more accurately described as the decade of agents.”
Andrej Karpathy, on the Dwarkesh Podcast dwarkesh.com →: Andrej Karpathy, on the Dwarkesh Podcast, 17 Oct 2025Archived →: copy of Andrej Karpathy, on the Dwarkesh Podcast, 17 Oct 2025 -
“NIST hosts technical convenings and conducts gap analyses to produce voluntary guidelines to inform industry-led standardization for AI agents.”
NIST, AI Agent Standards Initiative nist.gov →: NIST, AI Agent Standards Initiative, 17 Feb 2026Archived →: copy of NIST, AI Agent Standards Initiative, 17 Feb 2026
Not to be confused with
-
Agentic
Agentic is a property that comes in degrees, and many systems have some of it. An AI agent is one kind of system, in which the model chooses its own steps and tools to reach a goal.
-
Chatbot
A chatbot answers each message and then waits for the next. An AI agent takes a goal and carries out several actions on its own, such as searches or form entries, before reporting back.
-
Workflow
In a workflow, developers fix the order of model and tool calls in code, a form of orchestration. In an AI agent, the model chooses each next step as it goes. Anthropic separates the two this way.
-
Intelligent agent
The older textbook term covers any system that perceives its environment and acts on it, with or without a language model. In current use, an AI agent usually means a language model using tools toward a goal.
Questions
What is an AI agent?
An AI agent is software in which a language model works toward a goal by choosing its own steps. It calls tools such as a search engine or a code runner, reads each result and decides what to do next. It stops when it judges the task done or reaches a limit set by its developers. OpenAI’s Operator, which worked in its own web browser, was an early example.
What is the difference between an AI agent and agentic AI?
An AI agent is a kind of system. In it, a model pursues a goal by choosing its own steps and tools, usually in a loop. Agentic is a property that comes in degrees, so many systems are partly agentic without being agents. A fixed workflow that runs a model through steps set in code can be somewhat agentic. It is not an agent, because the model does not choose the steps.
How is an AI agent different from a chatbot?
A chatbot answers one message at a time and waits for the next. An AI agent takes a goal and acts on its own. It searches, clicks, edits or runs code, checks the results and keeps going. Because it acts, it needs safeguards a chatbot does not. OpenAI’s Operator, for example, returned logins and payments to the user and asked for approval before significant actions.
Do AI agents work reliably?
Tests in 2025 found clear limits on long tasks. In early 2025, METR found that leading models succeeded half the time at tasks that take a person about 50 minutes. That length had doubled about every seven months since 2019. In a Carnegie Mellon test, agents got office tasks wrong about 70% of the time. Gartner predicted in 2025 that more than 40% of agentic AI projects would be canceled by the end of 2027.
Where does the term AI agent come from?
By Alan Kay’s 1984 account, John McCarthy had the idea in the mid-1950s, and Oliver Selfridge coined the term agent a few years later. They pictured a program that pursues a goal and asks for advice when stuck. Software agents were a research field by the 1990s. The current sense spread in 2023, when BabyAGI and AutoGPT ran GPT-4 in loops toward goals set by users.
Watch
- LLM Agents MOOC | UC Berkeley CS294-196 Fall 2024 | LLM Agents: History & Overview by Shunyu YaoA UC Berkeley course lecture by Shunyu Yao of OpenAI on the history of LLM agents.
- Andrej Karpathy — “We’re summoning ghosts, not building animals”A long interview in which Karpathy explains why he calls this the decade of agents, not the year.
Cite this entry
“AI agent.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). AI agent. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “AI agent.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/