Definition:
An AI agent is a system that uses artificial intelligence to interpret information from its environment, select actions and carry them out to achieve a goal. It can operate as a software program or form part of a physical device, such as a robot. Its autonomy depends on the instructions, tools and permissions configured for it.
For example, an agent can check the status of an order and decide which check to perform next. Not all agents learn during use or need to operate without supervision: some follow stable rules, while others require approval before certain actions.
Contents
History of AI agents
The concept of an intelligent agent is part of artificial intelligence research into systems that perceive an environment and act on it. It includes everything from reactive agents that respond to specific situations to systems that plan actions or learn from experience.
Large language models (LLMs), which can process and generate text, have expanded the possibilities for creating software agents that can interpret requests, look up information and use tools. However, the concept is not limited to these models: it also applies to robotics and other decision-making systems.
How an AI agent works
How an agent works depends on its design, but it can be explained through a cycle of observation, decision and action:
- Receives information: interprets a request, a sensor reading or data from an application.
- Selects an action: uses its rules, its model or a planning process to decide the next step.
- Acts: uses a tool, generates a response or performs a permitted operation.
- Checks the result: uses the information obtained to continue, stop or ask for help.
In agents based on language models, this process can involve several tool calls. Anthropic’s guide to agents distinguishes these systems from workflows that follow a predefined sequence of steps.
A chatbot can be an agent’s interface, but not every chatbot is an agent. Answering questions through a conversation does not, by itself, mean selecting and executing a sequence of actions.
Types of AI agents
A common classification distinguishes agents by the information they use and how they choose their actions. These categories can be combined within the same system:
- Simple reflex agents: respond to the current situation using rules, without maintaining an internal history.
- Model-based agents: maintain a representation of the environment that helps them interpret what is happening.
- Goal-based agents: choose actions aimed at achieving a specific goal.
- Utility-based agents: compare possible outcomes according to a preference or value criterion.
- Learning agents: incorporate mechanisms to improve their behaviour through experience.
A hierarchical architecture organises decisions at different levels. A multi-agent system brings together several agents that interact, collaborate or compete. These are ways of organising a system, not capabilities that all agents possess.
Applications of AI agents
Agents can look up information, coordinate tasks or act on connected systems. For example, a customer service agent can check the status of an order, find out whether there is an issue and decide whether it can resolve the query or should refer it to a person.
To access this information, it can use integrations through APIs or protocols such as MCP. The actions available depend on its permissions: being able to look up an order does not mean it can modify it or issue a refund.
Agents can also help organise documents, review catalogues or research information. Their usefulness depends on the specific task, the available data and the controls in place.
Examples of AI agents
AI agents can specialise in software development, personal productivity, automation or system management. Current examples include:
- Hermes Agent: an open-source agent developed by Nous Research. It combines tools, persistent memory and reusable procedures and can run on a self-hosted computer or server.
- OpenClaw: an open-source platform for running and managing AI assistants on self-hosted infrastructure. It connects models with tools, files and communication channels.
- Grok Bot: an application that organises persistent agents capable of working with files, applications and websites from a shared cloud computer.
- Codex: OpenAI’s coding agent, which can explore repositories, modify files, run commands and verify the results of software development tasks.
- Claude Code: Anthropic’s agentic coding tool, which works with codebases, edits files, runs commands and integrates with terminals and development environments.
- Gemini Spark: Google’s personal AI agent, which can perform background tasks, use connected applications and manage workflows, skills and schedules under the user’s direction.
AI agent evaluation
A well-designed agent can reduce repetitive work and coordinate queries or actions that previously required several manual steps. It can also help maintain a consistent procedure and keep a record of the operations performed.
These benefits are not guaranteed. An agent can misinterpret a request, use incomplete data or compound errors. Calls to models and tools also add costs and execution time. When a task always follows the same steps, conventional automation may be simpler and more predictable.
How to implement an AI agent
Before using an agent, it is worth defining a specific task, the information sources and the limits on its actions. It is also necessary to establish how the correctness of the result will be checked.
An implementation can begin with read-only permissions and representative test cases. Sensitive actions, such as changing a price or sending a message, may require human approval. Logging operations, limiting attempts and defining when the agent should stop make it easier to detect and correct problems.
Saving information from a conversation is not the same as retraining the model. If the system incorporates memory or learning, it should specify what it retains, what it uses it for and how it is reviewed.