Definition
An AI agent is a software system that uses a large language model as its "reasoning engine" to pursue a goal across several steps, deciding what to do next and taking actions through tools. Instead of answering one prompt and stopping, an agent can break a request into subtasks, search the web, read files, query a database, fill in a form, send a message or run code, check the result and continue until the goal is met or it needs human input.
The key ingredients are a goal, a model that can plan, a set of tools the agent is allowed to use, memory of what it has done so far, and rules about when to stop or ask for approval. Products marketed as agents in 2025 and 2026 include OpenAI's agent features in ChatGPT, Microsoft Copilot agents, Google's Gemini agents, Salesforce Agentforce and many specialised tools for coding, customer service and sales.
Chatbot, assistant or agent?
- A chatbot responds to messages. You ask, it answers.
- An assistant is a chatbot with some tools — it may search the web or read a file you upload — but you drive each step.
- An agent receives a goal and decides the steps itself, calling tools in a loop with varying degrees of autonomy.
The line is blurry, and vendors use "agent" generously. A practical test: if the system can take an action in another application on your behalf without you approving each click, treat it as an agent and apply agent-level safeguards.
How an AI agent works
Most agents follow a loop often described as "plan, act, observe":
- Plan. The model reads the goal and decides the next step.
- Act. It calls a tool — for example, a calendar API or a web search.
- Observe. It reads the tool's result.
- Repeat or stop. It updates its plan, continues, or hands back to a human.
Frameworks and standards such as the Model Context Protocol (MCP), introduced by Anthropic in November 2024, make it easier to connect agents to business tools in a consistent, permissioned way.
Example
A recruiting coordinator is asked to schedule final-round interviews for six candidates with four interviewers across two time zones.
Without an agent, this takes most of a morning: checking calendars, emailing candidates, juggling replies. With an approved scheduling agent, the coordinator writes: "Schedule 45-minute final interviews next week for the six candidates in this list with Maria, Tom, Aisha and Ben. Each candidate meets two interviewers. Respect working hours in London and New York. Draft invitations but do not send them until I approve."
The agent reads the interviewers' availability, proposes a schedule that satisfies the constraints, drafts personalised emails and presents everything for review. The coordinator spots that one candidate requested a morning slot, asks the agent to adjust, approves the batch and the invitations go out. Total time: about 15 minutes. Crucially, the human set the boundaries and approved the irreversible action — sending emails to external people.
Where agents help at work
- Research: gathering and comparing information from many sources into a brief.
- Operations: triaging tickets, updating CRM records, reconciling data between systems.
- Software: writing, testing and fixing code.
- Customer service: resolving routine requests end-to-end and escalating complex ones.
- Personal productivity: inbox triage, meeting preparation, follow-up drafting.
Risks and how to manage them
Agents multiply both productivity and risk, because mistakes become actions rather than just words.
- Compounding errors. A wrong assumption early in the plan can cascade through many steps.
- Over-permissioned access. An agent with broad access to email, files and finance systems can do broad damage.
- Prompt injection. Malicious text on a web page or in an email can try to hijack an agent's instructions.
- Accountability gaps. Someone must own what the agent does.
Good practice: give agents the minimum permissions they need; require human approval for irreversible, external or financial actions; log every step; start with low-risk tasks; and test on realistic examples before scaling. Article 14 of the EU AI Act and the NIST AI Risk Management Framework both emphasise human oversight proportionate to risk.
Key takeaways
- An agent plans and executes multi-step tasks using tools, not just text.
- Autonomy is a dial: more autonomy means more need for guardrails.
- Keep humans approving actions that are irreversible, external or high-stakes.
- Start small, measure results and expand only where the agent proves reliable.