The first AI chat assistants were tools that produced a single answer to a question. The new generation of systems called “agents” can, when given a goal, plan the steps needed to reach it, use tools such as searching the web, reading and writing files, running code or navigating a website, and check the result, trying again if needed.
How do they work?
An agent usually works in a loop: it understands the task, makes a plan, uses a tool (for example running a search or opening a file), evaluates the result and decides on the next step. The loop continues until the task is done or the user’s approval is needed. What an agent can do depends on which tools and data it has been given access to.
Examples of use
- Research: Scanning dozens of sources and preparing a referenced report.
- Office work: Collecting information from emails into a table, scheduling meetings.
- Software: Finding and fixing a bug in a codebase, writing tests.
- Personal tasks: Comparing travel options, filling in forms (with approval for critical steps such as payment).
Risks
- Wrong actions: An agent may misunderstand an instruction and send an unwanted email or delete a file.
- Prompt injection: Malicious instructions hidden in a web page or document can try to steer the agent. This is one of the most debated security problems for agents.
- Excessive permissions: Broad access to sensitive areas such as bank accounts or company systems brings a risk of serious harm.
- Accountability: The user or organisation is ultimately responsible for actions the agent takes.
Principles for safe use
- Give the agent only the minimum permissions needed for the task.
- Set it to ask for approval before irreversible steps such as transferring money, purchasing, sending email and deleting.
- Review the log of actions the agent has taken.
- Be careful when letting an agent process documents and sites from unknown sources.
The difference between an AI agent and classic automation
The classic automation tools used in workplaces for years work according to pre-written rules, such as "If an invoice arrives in this folder, write the amount into the spreadsheet." When they meet a situation outside the rules, they usually stop. AI agents, by contrast, receive a goal in natural language, plan the steps themselves and can try a different route when something unexpected happens.
- Flexibility: An agent can try to understand a document whose format has changed; a rule-based tool cannot.
- Predictability: Classic automation gives the same result every time; an agent's behaviour can vary with the input, and it can make mistakes.
- Oversight: This is why keeping a record of every step and requiring human approval for important actions matter even more with agents.
Questions to ask before using an agent
The security community OWASP includes "excessive agency" in its list of the most important risks in large language model applications: if an agent is given more permissions than necessary, a misunderstood instruction or malicious content can cause real harm.
That is why it is useful to ask the following before starting to use an agent: Which accounts and files can it access? Do irreversible actions such as sending money, deleting or sending emails require my approval? Can I see every step it has taken afterwards? The US National Institute of Standards and Technology (NIST) AI Risk Management Framework similarly recommends continuous monitoring of the effects of AI systems and clearly defined responsibilities.
AI tools’ features, plans and prices change often. Always check the tool’s official page for current information.
Frequently asked questions
What is the difference between an agent and a chat assistant?
A chat assistant mainly produces answers; an agent takes actions by using tools towards a goal. As many assistants now include agent features, the line between them is increasingly blurred.
Sources and further reading
- OWASP, “Top 10 for LLM Applications” (prompt injection and excessive agency).
- NIST, “AI Risk Management Framework: Generative AI Profile” (NIST AI 600-1), 2024.
- Shunyu Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, ICLR, 2023.
- OWASP – Top 10 for LLM Applications 2025 (Excessive Agency)
- NIST – AI Risk Management Framework (AI RMF 1.0), 2023
Leave a comment
Your email address will not be published. Required fields are marked *