AI Glossary / Agents and automation / Beginner
Agent
Definition
In AI, an agent is anything that perceives its environment and takes actions in it to achieve a goal. A thermostat, a chess program and a warehouse robot all fit this description.
What it means
The word sounds human, but the idea is deliberately broad. An agent is defined by a loop: it receives information about the world, called percepts, and it does things, called actions. What sits between the two can be a simple rule or a sophisticated model.
This matters because it separates two questions people often merge. One question is what the system can sense and do. The other is how clever its decisions are. A system can be an agent with almost no intelligence, and a very intelligent model is not an agent until it can act on something.
A chatbot that only drafts text is not, by itself, much of an agent; a person still copies, checks and sends. Add the ability to file the draft, look up an order and issue a refund, and you have built an agent around the model, with real consequences attached.
The goal defines whether the agent is working well. A delivery robot's goal might be to deliver parcels quickly without breaking safety rules. Judging an agent means judging its outcomes in its environment, not its fluency in a demo.
For a business owner, the useful framing is: an agent is a worker with senses, hands and a job description. Before deploying one, you should know all three.
How it works
The classic picture has four parts. Sensors gather percepts from the environment. An internal decision process maps what has been perceived to a choice. Actuators carry out the action. A performance measure, given by the designer, says what counts as doing well.
The decision process can be a lookup table of rules, a learned model, a plan over many steps, or a mixture. Simple agents react to the current percept; more capable ones keep an internal picture of the world and plan ahead.
Modern AI agents often wrap a language model in a loop: the model reads the current state, chooses the next step such as calling a tool or asking for confirmation, observes the result and continues. The loop, the tools and the permission rules turn a text generator into an agent.
Because agents act, they need boundaries: which systems they may touch, what requires human approval, and how their actions are logged.
Three practical examples
- Home: a robot vacuum senses walls and dirt, and moves to cover the floor. It is an agent with a very small world.
- Operations: a support system reads a ticket, checks an order status and issues a standard refund within a set limit. The refund is a real action, so limits and approval rules matter.
- Analysis: a research assistant searches sources, reads pages and compiles a brief while a person reviews the result before it is used.
Think of it this way
Think of a night security guard in a building. The guard patrols, notices an open door, and either closes it or calls the manager, following standing instructions. The building is the environment, the patrol is the sensing, and the instructions are the decision process.
The analogy has limits. A guard understands context and can refuse a bad instruction. An AI agent follows its design and permissions, so the standing instructions have to carry the judgment the guard would supply.
Worked example
giving a drafting tool real hands, carefully
Suppose an online shop uses a chatbot to answer delivery questions. At first the bot only writes replies and a person sends them. It is software, but barely an agent.
The shop then connects the bot to the order system so it can check tracking and resend a confirmation email by itself. Now it perceives order data and acts on it. It has become a simple agent.
A sensible next step adds rules: refunds under a small amount can go through automatically; anything larger, unusual or repeated goes to a person. The rules, not the model's confidence, decide where the boundary sits.
Each expansion of what the agent may do is reviewed like a new hire's first week: more freedom only after the audited actions look right.
Illustrative case study
A fictional property manager automates maintenance requests. At first an AI reads tenant emails and drafts replies, with staff sending them. Complaints about slow answers persist.
The firm lets the system act: it may book approved contractors for standard jobs under a cost cap, and it must escalate anything involving safety, water or electricity. Every action is logged, and a weekly sample is reviewed.
Response times improve and staff spend their time on the escalated cases. The case is invented; the lesson is that value came from permissions and logging as much as from the model, because acting is what made the system useful and risky at once.
Common misconceptions and mistakes
- An agent is a robot. Most AI agents are pure software; the environment they act in is often email, files or a database.
- An agent understands its goal. It optimizes what its design measures, which may differ from what the owner actually wants.
- More autonomy is always better. Autonomy expands both usefulness and the cost of a mistake; the right level depends on how reversible the actions are.
Business use
Agents pay off where work follows a loop of reading, deciding and doing: triage, scheduling, reconciliation, monitoring and routine customer requests.
Adopt them by widening permissions in stages. Start read-only, then allow reversible low-value actions, then review logs before expanding. Name who owns the agent's mistakes, because the business, not the model, answers for them.
People also ask
Is a chatbot an agent?
Only loosely. It becomes a clearer agent when it can look things up and take actions, not just produce text.
Is an agent always learning?
No. Many agents follow fixed rules or fixed models; learning is a separate feature.
Who is responsible for an agent's actions?
The organization that deploys it. That is why permission boundaries and logs matter.
Check your understanding
A system drafts supplier emails but a person always checks and sends them. Is it acting as an agent yet, and what would change that?
Answer: Only weakly. It would act as a fuller agent if it could send emails or update records itself, which is exactly why permissions, limits and logging must be designed first.
Related terms
Learn next
AI Agents and Automation for Small Business covers guardrails, human approvals and cost controls for exactly this kind of system: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business. [C4]
More in Agents and automation
- Active learning agent - An active learning agent is a system that improves itself by choosing which examples it wants to be taught, instead of passively accepting whatever data it...
- Agent orchestration - Agent orchestration is the coordination of multiple AI agents or tool-using steps so they work together on a larger task: deciding which agent does what, in...
- Agent program - An agent program is the actual running software that implements an agent: the code that takes the current percepts and produces the next action.
- Agentic loop - An agentic loop is the repeating cycle at the heart of an AI agent: observe the current state, decide the next step, act, read the result, and repeat until...
- Automation - Automation is the use of technology to perform tasks with reduced human involvement, following defined rules or trained behavior.
- Batched automation - Batched automation is automation that processes work in grouped runs at scheduled times rather than handling each item the moment it arrives.
Sources
[S1] Russell and Norvig, AIMA glossary. [S5] Anthropic, Building effective agents. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Age progression Next: Agent experience