Guide
AI agents, explained simply
Everyone is talking about AI agents. The idea is simple: instead of just answering your questions, an agent takes a goal and does the steps to reach it, using apps and the web along the way. Here's how they work, in pictures, what they're already doing, the tools to try, and how to stay in control.
A chatbot answers. An agent does.
Chatbot
Find me a flight to London next Friday.
Here are five options with times and typical prices. You'll need to book on the airline's site.
Gives you information. You do the rest.
Agent
Book me a flight to London next Friday, morning, aisle seat.
- Searches airlines
- Compares times and prices
- Fills in the booking form
Found a 9:40 flight with an aisle seat. Shall I book it?
Does the steps for you, and checks before paying.
How an agent works: a simple loop
Every agent repeats four steps until the job is done. Step through a real example:
1. Goal: You say what you want done, not how to do it.
Example: "Book a table for 4 near the office on Friday at 8 pm, somewhere vegetarian-friendly."
2. Plan: The AI breaks the goal into steps.
Example: Find restaurants nearby → check menus for vegetarian options → check reviews → look for a free table at 8 pm.
3. Act: It uses tools: the web, apps, forms, email.
Example: Opens a booking site, searches the area, filters by rating and fills in the booking form.
4. Check: It looks at the result and decides what to do next.
Example: The first place is full at 8 pm, so it goes back to the plan and tries the next one. When it finds a table, it asks you to confirm before booking.
What an agent is made of
- BrainAn AI model that understands the goal, plans the steps and decides what to do.
- ToolsThe things it can use: web search, a browser, email, calendar, spreadsheets, other apps.
- MemoryWhat it keeps track of along the way: what it has tried, what worked, your preferences.
- GuardrailsThe rules on what it may do on its own and when it must stop and ask you.
How much freedom does it have?
- SuggestsTells you what to do; you do it.
- Most agents todayAsks before actingDoes the work, but checks with you before anything important.
- Acts, then reportsGets on with it and tells you afterwards.
- Fully on its ownRuns continuously with no human checks.
The leading agents stop and ask before anything important: OpenAI: ChatGPT agent asks for confirmation before consequential actions; Anthropic: Claude in Chrome stops before purchases and other sensitive actions.
What agents are doing in real life
- Customer serviceAgents answer customers, look up orders and process refunds. At Salesforce, AI agents now handle about half of customer conversations, with people taking the harder ones.Source: computing.co.uk
- Writing softwareCoding agents take a task, change many files and propose the change for a developer to review. At Anthropic, merged code changes rose from about 2,500 to over 30,000 a day in four months.Source: anthropic.com
- ResearchResearch agents read hundreds of web pages and write a sourced report. OpenAI says its deep research does in tens of minutes what would take a person many hours.Source: openai.com
- Everyday tasksAgents can plan and book travel, arrange dinner parties, rearrange meetings or turn a dashboard into slides, asking you before they pay or send anything.Source: openai.com
- IT and operationsAgents watch systems around the clock, investigate alerts and suggest the cause of an outage, so engineers start with an answer instead of a blank screen.Source: datadoghq.com
- Business workflowsSmall businesses connect agents to email, forms and spreadsheets to sort enquiries, follow up leads or prepare reports while they sleep, using tools like the ones below.Source: zapier.com
Agents you can use today
Ready-made agents inside apps you may already have.
- ChatGPT agentInside ChatGPT: browses, uses apps and completes tasks such as research, bookings and slideshows.Paid ChatGPT plans
- Claude in ChromeClaude reads the web page you are on, then clicks, types and fills in forms while you decide what happens.Paid Claude plans
- Deep researchGive it a question and it searches widely and writes a report with sources.In ChatGPT; similar features in Gemini and Claude
- Coding agentsGitHub Copilot, Claude Code and others take a programming task and propose finished changes.Free and paid plans
Build your own agents
No-code and low-code tools for connecting AI to your own apps and workflows.
- n8nA visual workflow builder for AI agents with 1,000+ integrations, popular with technical teams.Free self-hosted Community Edition; paid cloud plans
- Zapier AgentsCreate an AI teammate in plain English that works across 9,000+ apps.Free to start; paid plans
- MakeVisual, drag-and-drop automation with AI agents built in, connecting 3,500+ apps.See the site for current plans
- Microsoft Copilot StudioBuild agents that work across Microsoft 365, Teams and company data.Paid, usually through an employer
Use agents safely
What can go wrong
- Confidently doing the wrong thing. An agent can misunderstand the goal and act on it, fast. AI fails →
- Being tricked by hidden instructions. A web page or email can contain text that tells the agent to do something else. Prompt injection →
- Too much access. An agent with your email, files or payment details can do real damage if it goes wrong. AI and your privacy →
Three tips to start
- Start small. Try research, summaries or drafts first, where a mistake costs nothing.
- Keep "ask before acting" on. Let the agent prepare things, and approve anything that spends money, sends messages or deletes.
- Give the least access. Connect only the apps it needs, and never hand over passwords or card details in chat.
Keep exploring
For short definitions, see AI agent and agentic AI in the glossary. To see where agents fit in AI’s progress, see the AI ladder, where agents are level 3. For how agents are changing software teams, see how AI is changing the way software gets shipped.