An agent is the learning system that acts in an environment and receives feedback. In reinforcement learning, the agent observes the current situation, chooses an action, receives a reward or penalty and gradually learns which behaviour leads to better results.

In machine learning, the word agent is most often used in reinforcement learning. It describes the part of the system that makes decisions. The agent is not the whole world around it. It is the learner inside a task.

A simple agent can be a program learning to play a game. A more complex agent can be a robot learning to move, a trading system learning when to buy or sell, or an AI system learning how to complete a sequence of actions. The important point is always the same: the agent does something, the environment responds, and the agent uses that feedback to improve future behaviour.

What an agent means in reinforcement learning

In reinforcement learning, an agent is the system that tries to learn what to do. It does not receive a fixed list of correct answers for every possible situation. Instead, it learns through interaction.

The agent sees some information about the current state of the environment. Based on that information, it chooses an action. The environment then changes, and the agent receives feedback. That feedback is usually expressed as a reward, penalty or score.

Over many repeated interactions, the agent tries to learn a better strategy. In technical language, that strategy is called a policy.

A simple example of an agent

Imagine a small robot in a maze. The robot is the agent. The maze is the environment. At each step, the robot can move left, right, forward or backward.

If the robot gets closer to the exit, it may receive a positive reward. If it hits a wall, wastes time or moves away from the goal, it may receive a lower reward or a penalty. At first, the robot may move almost randomly. After many attempts, it can learn which actions usually lead to the exit.

This is the basic idea of an agent. It is not simply following a fixed script. It is learning from the consequences of its actions.

Agent, environment and feedback

An agent does not learn in isolation. It learns through a loop:

• Observation – the agent receives information about the current situation.
• Action – the agent chooses what to do.
• Environment response – the environment changes after the action.
• Feedback – the agent receives a reward, penalty or other signal.
• Learning – the agent updates its behaviour based on what happened.

This loop repeats many times. The agent gradually learns which actions are useful, which actions are risky and which actions should be avoided.

What the environment is

The environment is everything the agent interacts with. It can be a game, a simulation, a website, a robot control task, a market model, a recommendation system or a real physical space.

The agent does not control the whole environment. It only chooses actions. The environment then responds according to its own rules.

For example:

• in a chess program, the board and opponent are part of the environment,
• in a robot task, the room, objects and physics are part of the environment,
• in a marketing model, users, campaigns and conversions can be treated as part of the environment,
• in a game, the game rules, score and next screen state form the environment.

A well-designed environment is important because it shapes what the agent can learn.

What an action is

An action is something the agent can do. The set of possible actions depends on the task.

A game agent may choose to move, jump, turn or shoot. A robot may choose motor movements. A recommendation system may choose which product to show. A bidding system may choose how much to bid. A dialogue agent may choose what type of response or tool call to make.

The agent does not learn from thinking alone. It learns because actions create consequences. Without actions, there is no interaction and therefore no reinforcement learning loop.

What a state is

A state is a representation of the current situation. It tells the agent what is happening now.

In a game, the state may include the current screen, score, position of enemies and available moves. In a robot task, the state may include position, speed, sensor readings and object locations. In a business system, the state may include customer history, inventory, timing, price or campaign context.

The quality of the state representation matters. If the agent receives poor or incomplete information, it may learn the wrong behaviour. If it receives too much irrelevant information, learning can become slower or less reliable.

What a reward is

A reward is the feedback signal that tells the agent how good or bad an outcome was. It is often a number. Positive reward encourages behaviour. Negative reward or penalty discourages behaviour.

For example:

• a game agent may receive points for winning and penalties for losing,
• a robot may receive reward for reaching a target and penalty for falling,
• a delivery agent may receive reward for faster delivery and penalty for wasted movement,
• a recommendation system may receive reward when a user clicks, buys or stays engaged.

The agent usually tries to maximise cumulative reward, not only immediate reward. This is important because the best action now may not always produce the best long-term result.

Policy: how the agent decides

A policy is the agent’s strategy. It describes what action the agent should take in a given state.

A simple policy can be a fixed rule. For example: if the path ahead is blocked, turn right. A more advanced policy can be learned from experience and represented by a neural network, decision tree or another model.

The agent’s job is to improve its policy. At the beginning, the policy may be weak or random. After training, the policy should lead to better decisions more often.

Why the agent must balance exploration and exploitation

An agent has to solve a practical dilemma. Should it try something new, or should it use the action that already seems best?

Exploration means trying actions to learn more about the environment. It may produce mistakes, but it helps the agent discover better strategies.

Exploitation means using what the agent already knows. It can produce good results now, but it may prevent the agent from discovering an even better option.

For example, a recommendation agent may know that one product gets clicks. Exploitation means showing that product more often. Exploration means testing other products to learn whether they may work better for some users.

Agent vs model

The word agent is not the same as model.

A model is usually a mathematical system that makes predictions, classifications or decisions based on input data. An agent is a broader decision-making system that uses observations, actions and feedback.

An agent may contain one or more models. For example, an agent can use a neural network to decide which action to take. It can also use a model of the environment to predict what may happen next.

In practice, people often mix these words. But technically, the agent is the acting learner, while the model is one component that may help it decide.

Agent vs algorithm

An algorithm is the procedure used to learn or decide. An agent is the entity that applies that procedure in an environment.

For example, Q-learning is an algorithm. A game-playing system trained with Q-learning is an agent. Policy gradient methods are algorithms. A robot using a learned policy is an agent.

The distinction matters because the same algorithm can be used to train different agents in different environments.

Agent vs AI agent

Today, the term AI agent is also used more broadly. It can describe a software system that uses an AI model to plan tasks, call tools, search information, write code, use APIs or complete multi-step workflows.

That broader meaning is related, but not identical to the reinforcement learning meaning.

In reinforcement learning, an agent has a precise role: it acts in an environment and learns from feedback. In modern AI products, an AI agent may not always learn from rewards. It may simply follow instructions, use tools and execute a task with a language model.

For this reason, it is useful to ask: do we mean a reinforcement learning agent, or a general AI software agent?

Agent in large language model systems

In systems based on large language models, the word agent is often used when the system can do more than produce one text answer.

For example, an LLM-based agent may:

• read a user request,
• decide which steps are needed,
• search documents,
• call external tools,
• write or edit files,
• check intermediate results,
• return a final answer.

This does not automatically mean that the system is a reinforcement learning agent. It may not be learning from rewards during use. It may be an agent in the product sense: a system that can plan and act across several steps.

Agent and prompt engineering

Prompt engineering can influence how an AI agent behaves. A prompt can define the task, role, constraints, available tools, response format and safety rules.

However, a prompt is not the same as a reward signal. A prompt tells the system what to do. A reward signal evaluates the outcome of actions and can be used for learning.

This difference matters. A prompt can guide behaviour in one task. Reinforcement learning can adjust behaviour over many interactions based on feedback.

Agent and embeddings

Embeddings can support agent behaviour in systems that need retrieval, memory or similarity search.

For example, an AI agent may use embeddings to find relevant documents before answering a question. It may search a knowledge base, compare a user request with previous cases or retrieve examples that help it complete the task.

In this setup, embeddings do not make the agent learn by themselves. They help the agent access relevant information. The agent still needs a decision process that determines what to search, what to use and what action to take next.

Agent and human feedback

Feedback does not always have to come from a simple numerical score. In some systems, feedback can come from people.

For example, a human may compare two outputs and say which one is better. That preference can be used to train a reward model. The agent or model can then be adjusted to produce outputs that better match human preferences.

This idea is important in reinforcement learning from human feedback. It is one reason why the term agent appears in discussions about AI alignment, reward models and safer model behaviour.

Agent and reward hacking

Reward hacking happens when an agent learns to optimise the reward signal while missing the real purpose of the task.

This is a risk because the agent does not automatically understand human intention. It optimises what the reward function measures. If the reward is incomplete, the agent may find a shortcut.

For example, if a cleaning robot receives reward only for moving over dirty areas, it might learn to create or spread dirt if the environment allows it. If a recommendation system is rewarded only for clicks, it may learn to show sensational content that gets clicks but harms user trust.

The problem is not that the agent is malicious. The problem is that the feedback signal is badly specified.

Agent and environment design

A reinforcement learning agent depends heavily on the environment used for training.

If the environment is too simple, the agent may learn behaviour that does not transfer to real use. If the environment is unrealistic, the agent may optimise for artificial conditions. If the environment lacks important constraints, the agent may learn actions that would not be acceptable in practice.

This is why simulations must be designed carefully. A robot trained in simulation may behave differently in the physical world. A business agent trained on historical data may fail when market behaviour changes. A game agent may exploit a bug in the game mechanics instead of learning the intended strategy.

Agent and delayed rewards

Some tasks have immediate feedback. Others have delayed feedback.

In a maze, the agent may not know whether a move was good until many steps later. In marketing, a campaign decision may not show its result until days or weeks later. In logistics, a route decision may affect later delivery times.

Delayed rewards make learning harder. The agent must connect earlier actions with later outcomes. This is called the credit assignment problem: which past action deserves credit or blame for the final result?

Single-agent and multi-agent systems

Some tasks involve one agent. Others involve multiple agents.

In a single-agent system, one learner interacts with the environment. For example, one robot learns to move through a room.

In a multi-agent system, several agents interact with the same environment and sometimes with each other. They may cooperate, compete or do both. Examples include game-playing agents, trading bots, traffic control systems, swarm robotics or multi-player simulations.

Multi-agent systems are harder because the environment changes as other agents learn and act. What works today may stop working when another agent changes its strategy.

Where agents are used

Agents are useful in tasks where decisions are sequential. One action affects the next situation, and the best outcome depends on a series of choices.

Common areas include:

• Games – learning strategies through repeated play.
• Robotics – learning movement, navigation or manipulation.
• Recommendation systems – choosing what content, product or offer to show next.
• Advertising and bidding – adjusting bids or budgets based on performance feedback.
• Logistics – planning routes, stock movement or resource allocation.
• Energy systems – controlling storage, demand response or optimisation decisions.
• Conversational AI – managing multi-step interaction, tool use or dialogue strategy.
• Simulation training – testing decision strategies before real deployment.

Common mistakes when explaining agents

The word agent is often used too loosely. This can make the concept confusing.

Common mistakes include:

• calling every AI model an agent – a passive model that only predicts is not necessarily an agent,
• forgetting the environment – an agent needs something to act in,
• confusing feedback with instruction – a prompt tells the system what to do, while feedback evaluates what happened,
• assuming the reward equals the real goal – a reward is only a proxy for the goal,
• ignoring delayed effects – the best immediate action may not be the best long-term action,
• treating tool use as learning – an AI system can use tools without learning from reward,
• overstating autonomy – many AI agents still operate under strict human-defined rules and constraints.

Why the agent concept matters

The concept of an agent is useful because it shifts the focus from static prediction to action.

A normal prediction model answers a question such as: what is likely to happen? An agent asks a different question: what should I do next?

That difference is important. Acting creates consequences. Consequences create feedback. Feedback changes future behaviour. This is why agents are central to reinforcement learning and many discussions about autonomous AI systems.

How to remember agent

An agent can be compared to a learner inside a game. It sees the current situation, makes a move, receives a score and slowly improves.

The agent is not the game itself. It is the player learning how to play.

Related terms

• Machine learning – the broader field in which systems learn patterns from data and use them for prediction, classification or decision support.
• Reinforcement learning – a machine learning approach where an agent learns through actions, environment feedback and rewards.
• Environment – the world, task or simulation the agent interacts with.
• State – the current situation observed by the agent.
• Action – a choice the agent can make in the environment.
• Reward – feedback that tells the agent whether an action or outcome was good or bad.
• Policy – the strategy that maps states to actions.
• Exploration – trying new actions to discover better behaviour.
• Exploitation – using the best-known action based on current knowledge.
• Reward hacking – a situation where an agent optimises the reward signal while missing the real purpose of the task.
• RLHF – reinforcement learning from human feedback, where human preferences are used to shape model behaviour.
• Large language model (LLM) – a language-focused AI model that may be used as part of a broader agentic system.
• Prompt engineering – the practice of designing prompts that guide language model outputs and tool-using AI systems.
• Embedding – a numerical representation of content that can help agents retrieve or compare information.
• Bagging – an ensemble method in machine learning; not an agent concept, but related to model training and prediction stability.

Sources and further reading

• What is reinforcement learning? – cloud.google.com – June 2026 – explains reinforcement learning as a process where an agent learns behaviour through interaction with an environment and feedback.
• Part 1: Key Concepts in RL – spinningup.openai.com – June 2026 – introduces agents, trial-and-error learning, reward and punishment in reinforcement learning.
• Part 2: Kinds of RL Algorithms – spinningup.openai.com – June 2026 – explains model-based and model-free reinforcement learning and how agents can use models of the environment.
• What is reinforcement learning? – ibm.com – June 2026 – describes reinforcement learning concepts including actions, states, rewards and decision-making.
• What is Reinforcement Learning? – aws.amazon.com – June 2026 – gives a business-oriented explanation of reinforcement learning and trial-and-error optimisation.
• Learning through human feedback – deepmind.google – June 2026 – explains how human preferences can be used as feedback for training agents through reward prediction.