Examples of Intelligent Agents in AI: The Five Types, Real-World Examples, and How Agent Memory Fits In

What is an intelligent agent in AI? An intelligent agent is a system that perceives its environment, decides what to do, and acts toward a goal. The best examples of intelligent agents in AI are the ones you can already point to: a thermostat, a robot vacuum, a self-driving car, a trading system, AlphaGo, and the coding agents many of us now use daily. The framing comes from Stuart Russell and Peter Norvig's textbook Artificial Intelligence: A Modern Approach (4th edition, Pearson, 2020), which defines an agent as anything that perceives its environment through sensors and acts on it through actuators. The loop is perceive, decide, act, repeat. A "rational" agent picks the action expected to do best by some measure of success. That is a looser definition than most people expect. A chatbot qualifies. So does a light switch with a timer, though a weak one. The five types of intelligent agents, with an example of each Russell and Norvig sort agents by how much they do between perceiving and acting. The same five categories appear in the 4th edition. Here they are from simplest to most capable. 1. Simple reflex agents: the thermostat A thermostat reads the temperature and fires a rule: below 19°C, heat on. It has no history and no model of the room. That is fine, because the job is narrow and the environment is easy to read. 2. Model-based reflex agents: the robot vacuum A robot vacuum can't see the whole apartment at once. It keeps an internal map of what it has cleaned and where the furniture is, and it acts on that map as well as on its current sensor reading. 3. Goal-based agents: navigation and self-driving A navigation system has a destination. It compares possible routes against that goal and picks one. A self-driving stack does the same at a much harder scale. The SAE J3016 standard defines six driving-automation levels, from 0 (no automation) to 5 (full automation), which is a useful yardstick for how much of the goal-seeking the machine owns. 4. Utility-based agents: trading and recommendation systems When several goals compete, you need a way to score trade-offs. A trading system weighs expected return against risk. A recommender weighs relevance against novelty. The agent maximizes a utility function, not a yes/no goal. 5. Learning agents: AlphaGo and reinforcement learners Learning agents improve with experience. DeepMind's AlphaGo beat Lee Sedol 4-1 in Seoul in March 2016 (DeepMind, as reported in the Nature paper and match coverage). It learned from human games and then from playing itself. This is the category that most modern AI hype points at. Best examples of intelligent agents in use today The textbook types are a good vocabulary, but real systems mix them. Here is how the common examples of intelligent agents compare. The ratings are our own judgment, not a benchmark. ExampleAutonomyEnvironmentLearning Autonomous vehiclesHigh (SAE levels 3-5)Open, fast-changing roadOffline training, fleet updates Virtual assistantsLow to mediumPhone, speaker, appsMostly model updates Customer service agentsMedium, with human handoffTickets, chat, knowledge baseTuned on past tickets Coding agents (Claude Code, Copilot agents)Medium, user approves actionsRepo, terminal, toolsLittle by default; memory is bolted on Robotics (warehouse, surgical)Varies widelyPhysical, structuredOften trained in simulation Fraud detectionHigh on scoring, low on final actionTransaction streamsRetrained on labeled fraud Multi-agent systemsDepends on the rolesShared task or marketVaries per agent A few notes on the table. Customer service agents from vendors such as Aisera sit in the middle on autonomy because most deployments hand hard cases to a person. Fraud systems score nearly every transaction on their own, but a human often makes the call on the edge cases. Multi-agent systems are less a type of agent than a way of arranging several of them, such as a planner, a coder and a reviewer passing work along. Coding agents deserve a closer look because many readers here use them. Claude Code, Cursor and Codex read your repo, run commands and edit files. They're capable. They also start every session knowing nothing about you. Why LLM agents forget: the memory problem A language model has no state between calls. Whatever it "remembers" is whatever fits in the context window of the current session. When the chat ends, it's gone. Agent designers usually describe four kinds of memory, borrowed loosely from cognitive science: Short-term (working) memory: the current context window. Episodic memory: what happened, such as "we tried Redis for the job queue and it was overkill." Semantic memory: stable facts, such as "this project uses Postgres 16." Procedural memory: how you like things done, such as "always use pnpm, never npm." Most agents ship with only the first. Everything else has to be stored outside the model and fed back in, within a token budget. Paste too much and you burn your limits. Paste too little and the agent repeats old mistakes. Where Open Agentic Brain & Memory fits At Open Agentic Brain & Memory (OpenAB), we think of memory as a layer that belongs to you, not to any one agent. OpenAB is free, open-source (Apache-2.0), self-hosted memory shared across Claude Code, Cursor, Codex, Antigravity and Kilo Code (beta). Decide something in Claude Code on Monday and Cursor already knows it on Tuesday. Here is what it does. It captures decisions, fixes and preferences from your sessions and tags them by project, found by walking up to the nearest .git. They're stored in a local SQLite database. At session start, a hook injects a briefing of at most 2,000 tokens, ready in 84 ms (p50). Tools can also recall on demand over MCP (brain_recall, brain_remember, brain_context), which takes 28 ms (p50) across 10,000 memories. A local control panel lets you read, edit, pin or delete any memory. Some numbers, with their source: we measured these ourselves. Loading context for a 10,000-memory store takes 1,325 tokens instead of 1,131,367. On the LoCoMo long-conversation benchmark, OpenAB keeps 97% of the full-conversation score while reading 18% of the tokens. That comparison used the same answering and judging model (gpt-6-luna) over all 1,540 questions. It's our own run, so run brain eval on your own history before you trust it. For sync, memory is age-encrypted, the key stays in your OS Keychain, and the blob lives as a Release asset in your own private GitHub repo. There is no account and no server of ours. The trade-off is plain: only you hold the key, and if you lose it, nobody can recover your memory. What it does not do: it is not an agent. It doesn't plan, call tools or write code. It's the memory those agents read from. Hosted memory APIs such as Mem0 (which also offers an open-source option) take a different approach, and you should check their current docs for specifics. Ours is local-first and owner-held. You can see how it works on the OpenAB site. How to choose or build an intelligent agent: a checklist Whether you're buying an agent or building one, these five questions do most of the work. Task scope. Can you describe the job in one sentence? Narrow agents fail less often. Tool access. What can it touch? Give it the least access that still gets the job done. Memory. What does it remember between sessions, where is that stored, and can you inspect and delete it? Evaluation. How will you know it got better or worse? Test on your own tasks, not only on vendor benchmarks. Safety and human oversight. Decide which actions need approval before you turn it loose, not after the first incident. Intelligent agents range from a thermostat to AlphaGo, and the ones you'll use most today are LLM agents that act well and remember poorly. Fix the memory and most of the daily friction goes away. If that's your friction, Install Open Agentic Brain & Memory (OpenAB). FAQ What are the best examples of intelligent agents in AI? Common examples are thermostats, robot vacuums, navigation and self-driving systems, trading and recommendation engines, AlphaGo, virtual assistants, customer service agents and coding agents. They differ in autonomy, environment and how much they learn. What are the five types of intelligent agents in AI? Russell and Norvig's five types are simple reflex, model-based reflex, goal-based, utility-based and learning agents. Each adds something to the one before: memory of the world, a goal, a way to score trade-offs, and finally the ability to improve. Is ChatGPT or Claude an intelligent agent? On their own, they're language models that answer prompts. Wrapped in a product that lets them use tools and take actions, such as Claude Code, they act as agents. The model is the brain, and the agent is the loop around it. Is a thermostat an intelligent agent? By the textbook definition, yes. It senses temperature and acts on it. It's a simple reflex agent, so it isn't intelligent in the everyday sense. What is the difference between an AI agent and an intelligent agent? The terms mostly overlap. "Intelligent agent" is the older textbook term and covers anything that perceives and acts. "AI agent" is how people usually refer to LLM-driven systems that plan and use tools. How do intelligent agents remember information between sessions? They don't, unless something outside the model stores it. That can be a file, a database or a memory layer that writes relevant notes back into context at the start of the next session. Is Open Agentic Brain & Memory an AI agent? No. OpenAB is a memory layer. It stores and retrieves what your coding agents learn, so that Claude Code, Cursor, Codex and others share it. The agents still do the work.