What is an AI agent?

How AI agents work

At its core an AI agent runs a loop. It perceives the current situation (a user request plus any relevant context or tool output), reasons about the next step using a large language model, acts by invoking a tool or producing an answer, and then observes the outcome before deciding what to do next. Amazon Web Services describes agents as autonomous entities that perceive their environment, reason, and take actions to achieve defined goals, moving through the stages of perceive, reason, act, and learn.

Google Cloud frames the same idea in terms of components. A reasoning model such as Gemini or Claude acts as the engine that thinks; tools give the agent the ability to interact with the outside world through APIs, databases, and search; memory holds short-term working state and longer-term knowledge; and an orchestration layer manages the plan, tracks task status, and decides when to call which tool.

In practice, this means an agent does not just generate text. Asked to 'find the three cheapest flights next Friday and draft an email', it will break the goal into steps, call a flight-search tool, read the results, compare them, and only then write the email, looping through reason and act until the task is complete.

Agentic AI versus a chatbot

A conventional chatbot, including a plain large language model, responds to a prompt and stops. It has no built-in ability to take actions in the world, hold a persistent goal across many steps, or decide on its own what to do next. Each turn is largely independent.

Agentic AI adds autonomy and tools on top of that reasoning core. Google Cloud defines AI agents as software that pursues goals and completes tasks on behalf of users, showing reasoning, planning, and memory with enough autonomy to make decisions, learn, and adapt. The practical dividing line is action: an agent can book the meeting, run the query, or edit the file, whereas a chatbot can only tell you how to do it.

The distinction is a spectrum rather than a hard boundary. A model that can call a single calculator sits near the chatbot end; a system that plans a multi-step research task, spawns sub-tasks, and self-corrects sits firmly at the agentic end.

The ReAct pattern

Much of modern agent design traces back to ReAct, short for 'Reasoning and Acting', introduced by Shunyu Yao and colleagues in a 2022 paper (arXiv:2210.03629). ReAct prompts a language model to interleave reasoning traces ('thoughts') with task-specific actions in a single loop, rather than treating thinking and acting as separate stages.

The authors argue that reasoning traces help the model induce, track, and update its plan and handle exceptions, while actions let it pull real information from external sources such as a knowledge base or environment. By checking its reasoning against a live tool, for example a Wikipedia search API, ReAct reduces the hallucination and error propagation that pure chain-of-thought reasoning is prone to.

This think-act-observe template is now the backbone of popular agent frameworks such as LangChain and LangGraph, and of the tool-use loops built into products like Claude and ChatGPT. When you see an agent 'show its work' before calling a tool, you are usually looking at a descendant of ReAct.

Tools, memory, and planning

Tools are what turn a language model into an agent. A tool is a function the model can call, described by a name, a purpose, and an input schema, such as a web search, a SQL query, a code interpreter, or a payment API. The model decides when to call a tool and with what arguments; the surrounding system executes it and feeds the result back. Standardising how models discover and call these tools is the specific problem addressed by the Model Context Protocol.

Memory lets an agent carry context beyond a single step. Short-term memory holds the current task's working state, while long-term memory, often stored in a vector database, lets the agent recall earlier interactions, documents, or facts. Planning is the third pillar: the agent decomposes a goal into ordered sub-steps, sometimes revising the plan as new information arrives.

Together these capabilities are what separate a genuine agent from a single model call. The quality of an agent depends less on raw model intelligence and more on how well its tools, memory, and planning are wired together.

Multi-agent systems

Complex work is increasingly handled not by one agent but by several that collaborate, an arrangement known as a multi-agent system. A common design uses an orchestrator or 'lead' agent that decomposes a task and delegates sub-tasks to specialised worker agents, then synthesises their outputs.

The appeal is division of labour and parallelism: a research task might spawn separate agents to gather sources, verify facts, and draft prose, each with its own tools and focus. Anthropic and others have reported that such architectures can outperform a single agent on broad, open-ended tasks, though at higher token cost and coordination complexity.

Multi-agent designs also raise new questions of communication and control, which has prompted emerging interoperability standards for how agents talk to each other and to shared tools.

Real-world use cases

Coding is the most mature application. Agents such as Claude Code and other 'agentic' developer tools read a codebase, plan a change, edit files, run tests, and iterate until the task passes, all with minimal step-by-step instruction. Customer support agents resolve tickets by looking up account data and taking actions in back-end systems rather than just suggesting replies.

Beyond software, agents power research assistants that browse and synthesise sources, operations bots that reconcile data across enterprise systems, and personal assistants that, given the right connectors, can manage a calendar or inbox. OpenAI's ChatGPT agent and similar products let a model carry out multi-step tasks in a browser or connected apps on the user's behalf.

In every case the pattern is the same: the value comes from the agent completing a workflow end to end, not from producing a single answer.

Limitations and risks

Agents inherit every weakness of the underlying model, including hallucination, and compound them over multiple steps. An early mistake can propagate through a long chain of actions, so reliability on long-horizon tasks remains a genuine constraint rather than a solved problem.

Autonomy also introduces safety and security risks. An agent that can take real actions can take the wrong ones, and because it reads untrusted content and calls external tools, it is exposed to prompt-injection attacks that try to hijack its behaviour. This is why production agents run with scoped permissions, human-in-the-loop approval for consequential steps, logging, and hard limits on what they can touch.

The consensus across vendor guidance from AWS, Google Cloud, and others is that agents are best deployed with clear guardrails and oversight. They extend what a model can do, but responsibility for the outcome stays with the humans who deploy them.

Frequently asked questions

What's the difference between an AI agent and a chatbot?

A chatbot responds to prompts one turn at a time and cannot act in the world. An AI agent adds autonomy, tools, memory, and planning so it can pursue a goal over many steps and take actions on your behalf, such as running a query or editing a file. Google Cloud defines agents as software that completes tasks for users with reasoning, planning, and memory, which is precisely what a plain chatbot lacks.

What is agentic AI?

'Agentic AI' is the broad term for AI systems that behave as agents: they have enough autonomy to plan and complete multi-step tasks rather than just answering questions. AWS defines agentic AI in terms of autonomous entities that perceive, reason, act, and learn to achieve goals. It describes a capability level, not a specific product.

What is the ReAct pattern in AI agents?

ReAct ('Reasoning and Acting'), introduced by Yao et al. in 2022 (arXiv:2210.03629), is a method that interleaves a model's reasoning traces with tool actions in one loop. Reasoning helps the model plan and handle exceptions, while actions let it fetch real data, which the paper shows reduces the hallucination seen in reasoning-only approaches. It underpins most modern agent frameworks.

Do AI agents use large language models?

Most modern AI agents are built around a large language model that serves as the reasoning engine, for example Gemini, Claude, or GPT models. The model decides what to do and which tools to call; the surrounding orchestration, tools, and memory turn those decisions into real actions. The term 'agent' predates LLMs, but today it almost always refers to LLM-powered systems.

Are AI agents reliable enough to run without supervision?

Not for consequential work. Agents compound the errors of their underlying model over many steps and are vulnerable to prompt-injection attacks, so vendor guidance from AWS, Google Cloud, and others recommends scoped permissions, human approval for high-impact actions, and logging. They are powerful assistants, but responsibility for outcomes stays with the humans deploying them.

What is a multi-agent system?

A multi-agent system uses several specialised agents that collaborate, often with an orchestrator agent delegating sub-tasks to workers and combining their results. It enables division of labour and parallelism on broad tasks, at the cost of higher token usage and coordination complexity.

Related

Explainers: MCP, LLM

Glossary: Agentic AI, Large language model (LLM), Multi-agent system

Sources

• Google Cloud — What are AI agents?
• AWS — What is Agentic AI?
• IBM Think — What are AI agents?
• Yao et al. — ReAct: Synergizing Reasoning and Acting in Language Models (arXiv:2210.03629)
• Anthropic — Model Context Protocol