AI agent
- Agents
- Standard term
An AI agent is a system that uses a model to choose and carry out steps toward a goal, using tools such as search or email. It observes results until completion, a limit, failure, or a handoff.
The model proposes a next action, and the surrounding software decides whether an available tool may carry it out. The result returns as information for the model's next decision. This loop can adapt to a failed step, but it can also stop or repeat an unhelpful action.
Where you’ll meet it
You ask your assistant app to move your dentist appointment to next week. While you're in a meeting, it checks your calendar, emails the office, reads their reply, and moves the event. The calendar access it was given would also let it delete any event you have.
What to do
Check the app's actual calendar and email permissions. Where supported, limit it to the appointment involved or require approval for sending and changing events. If access remains broad, use a workflow that prepares proposed changes for you before they take effect.
Easy to mix up
A chatbot is the conversation you type into, and many chatbots now have agents inside them. An AI agent is the system behind the work, which chooses and carries out steps toward a goal.
Check yourself
Your support agent can look up orders and issue refunds. Which protection more reliably prevents a refund over $100?
Go deeperA single step can do harm, and retries can repeat it
In AI research, an agent has long meant a system that senses its surroundings and acts on them, as in Russell and Norvig's textbook Artificial Intelligence: A Modern Approach. Early language-model agents such as ReAct wrote tool requests as plain text. Function calling instead supplies structured requests with a tool name and inputs.
The number of steps and the possible harm are separate questions. A single model request can trigger a payment, while a long loop can stay read-only. OWASP's risks for language-model applications include excessive agency: excessive functionality, permissions, or autonomy. Controls should match the actual effects the system can cause.
A tool may complete an action even if its confirmation is lost, so retrying can repeat the effect. An idempotent operation has no extra effect when the same request is repeated. A common protection is a unique action key that lets the receiving service recognize and reject duplicates.
Sources
- LLM06:2025 Excessive Agency, OWASP GenAI Security Project
- Function calling and other API updates, OpenAI
- Artificial Intelligence: A Modern Approach, Agents and Environments, aima.cs.berkeley.edu
- ReAct original author repository and prompt, raw.githubusercontent.com
- Stripe API: Idempotent requests, docs.stripe.com