How AI agents actually work (and why they're nothing like chatbots)

AI agents aren't chatbots with extra steps. The difference is bigger than most people realise, and it matters now because 2026 is the year agents stopped being a demo and started doing real work.

Here's what's actually happening under the hood.

Chatbots vs agents: the core difference

A chatbot answers one message at a time. You ask, it replies, the conversation ends. Next time you open it, it has forgotten everything. It has no goals beyond responding to whatever you just typed.

An AI agent is different in every one of those dimensions. It maintains state across steps, sets and pursues a goal, uses external tools to take real-world actions, and keeps going until the task is complete or it hits a wall it can't get past. You don't send it a message and wait for a reply. You hand it a task and it works through it.

A simple example: ask ChatGPT "what's the weather in Tokyo?" and it tells you. Give an agent the task "book me the cheapest flight to Tokyo next weekend and add it to my calendar" and it searches flight APIs, compares prices, picks one, books it, and creates the calendar event. That's the gap.

The five components every agent has

Perception

The agent needs to see something to act on it. Modern agents can read text, browse websites, look at screenshots, parse emails, pull data from APIs, and scan files. Whatever the task requires, perception is how the agent gets the raw material.

Reasoning

This is the model doing the hard cognitive work. Given what it can see, what should it do next? The reasoning layer breaks a goal into sub-steps, evaluates options, and decides which tool to call or which action to take. In 2026, this is still the part that fails most often, because reasoning over long chains of steps compounds errors.

Memory

Agents have two kinds. Short-term working memory holds everything from the current task: what's been done, what failed, what the user said. Long-term memory persists across tasks, so the agent can remember your preferences, past projects, or recurring patterns. Without long-term memory, every task starts from zero.

Planning

Before acting, a good agent maps the path from the current state to the goal. It breaks the job into ordered sub-tasks, estimates what could go wrong, and adjusts the plan as new information comes in. This is what separates an agent from a glorified macro.

Action

This is the agent actually doing things: running code, sending HTTP requests, writing files, clicking UI elements, calling APIs. The set of actions available to an agent is its "tool belt". Give it access to your email and it can send messages. Give it a browser and it can scrape or fill forms. The tools define what the agent can actually change in the world.

Why one agent is rarely enough

For simple tasks, a single agent works fine. For anything complex, you need a network. The reason is specialisation and reliability.

A customer support pipeline is a clear example. A classifier agent reads the incoming complaint and decides its category. A lookup agent checks the order database. A writing agent drafts the response using what the lookup found. A quality agent reviews the draft before it goes out. No single model does all of this as well as four specialists handing off to each other.

This is how Anthropic's Claude Agent SDK works in practice: you define specialised agents, give each a set of tools, and wire them together into a pipeline. The same architecture powers products like Cursor, the AI coding assistant that SpaceX acquired for $60 billion in June 2026. Cursor runs multiple agents in parallel: one reads your codebase, one writes new code, one runs tests and reports back.

The frontier labs all have their own takes on how to do this. Anthropic's Claude agents, OpenAI's operator mode in ChatGPT, and Google's Gemini agents are the three main options in production today. Each has different tool access, different memory architectures, and different strengths.

What agents still can't do well

This matters as much as what they can do.

They hallucinate. An agent reasoning through 30 steps has 30 chances to make something up and act on it. Compounding errors are a real problem, especially when the agent has tool access and can take actions that are hard to reverse.

Very long-horizon tasks break down. Agents are good at tasks that take minutes to hours. Tasks that require weeks of ongoing judgment, adapting to genuinely novel situations, or making high-stakes calls they haven't encountered before are still hit-and-miss.

Guardrails are non-negotiable. An agent with broad tool access and no constraints can do a lot of damage quickly. Rate limits, approval steps for high-risk actions, and sandboxing are not optional. Research into AI trust consistently finds that people are still far from comfortable letting AI systems take autonomous action. That caution is earned.

They're expensive when things go wrong. A chatbot that gives a bad answer costs you a few seconds. An agent that makes 40 API calls down the wrong path before failing can cost real money, or take real actions you need to undo.

The bottom line

The shift from chatbots to agents is the most meaningful change in how AI is actually used since the GPT moment in 2022. Chatbots made information retrieval conversational. Agents make task execution autonomous.

That is a much bigger deal, with proportionally bigger risks. The Transformer architecture that Noam Shazeer co-invented in 2017 powers every agent running today. What he and his co-authors built as a research paper is now making hiring decisions, writing production code, and booking flights.

Understanding the five components, the multi-agent model, and where agents genuinely fall short puts you in a better position to know when to trust them, when to add a human check, and when to just use a chatbot.