The short answer
An AI agent is software built on a large language model that plans and carries out steps towards a goal on your behalf, using tools such as a browser, your files or connected apps, instead of only replying in chat. Google's Machine Learning Glossary defines it as software that can reason about user inputs to plan and execute actions for the user. In November 2026, examples include ChatGPT Work and dots, Claude's Cowork tasks, Microsoft Copilot Cowork and Perplexity Computer.
What an AI agent is, in plain English
A chatbot answers. An agent acts. You give an agent a goal, such as "find the five most recent pricing changes from our three main competitors and put them in a spreadsheet", and it works out the steps itself: search, open pages, copy figures, build the file, and hand it back.
Google's Machine Learning Glossary describes an agent as software that can reason about user inputs in order to plan and execute actions on behalf of the user. Anthropic, which makes Claude, draws a useful line in its engineering guide Building effective agents (December 2024). It calls systems where LLMs and tools follow predefined code paths workflows, and systems where the LLM dynamically directs its own process and tool use agents. Both count as agentic systems.
Under the hood an agent is a loop. A large language model looks at the goal, chooses a tool, reads the result, decides the next step and repeats until it is done or needs you. Every step adds text to its context window, which is why long agent runs can lose track of earlier details.
AI agents you can use today, November 2026
Every major assistant now has an agent mode. They differ in where they run and what they are allowed to touch.
| Agent | Maker | What it does and who gets it |
|---|---|---|
| ChatGPT Work | OpenAI | Multi-step tasks that produce documents, presentations and spreadsheets, use connected apps and work in a browser. Limited on Free and Go (desktop app only); Plus and Pro on desktop, web and mobile. |
| Dots | OpenAI | Always-on agents with their own cloud computer, powered by GPT-6 Astra. Eligible Pro and Business Premium users aged 18+; Pro access excluded the EEA, Switzerland and the UK at launch. |
| Cowork in every conversation | Anthropic | Claude decides whether to answer or take on a task and returns editable files; cloud and scheduled tasks keep running with your computer off. Rolling out to Pro, Max and Ultra. |
| Copilot Cowork and Autopilot | Microsoft | Cowork takes multi-step work and returns a finished result. Autopilot, previously called Scout, is an always-on cloud agent with its own account, mailbox and files, rolling out to Frontier in November 2026. |
| Tasks in Gemini | A toggle that switches Gemini from chat to tasks so it keeps working across Google apps while you are away; announced on 13 October 2026 for Google AI Pro and Ultra subscribers. | |
| Perplexity Computer | Perplexity | Runs research, documents, presentations and scheduled tasks in a cloud sandbox, paid for with credits. |
| Vibe Work | Mistral AI | Breaks the job into steps, shows its progress and tool calls, and asks for confirmation before sensitive actions such as sending email, posting messages or changing data in external tools. |
For a hands-on walkthrough of these modes, see how to use AI agents and how to automate tasks with AI.
Why AI agents matter at work
Agents change what you can hand over. A chatbot saves you writing time; an agent can save you the clicking between apps that surrounds the writing. Good agent jobs share three traits: many small steps, a clear finished output, and low cost if something goes slightly wrong. Examples are compiling a research brief from public sources, turning a folder of notes into a first-draft deck, or checking a list of supplier websites for updated terms.
Anthropic's own advice is worth repeating to any team: find the simplest solution that works and add complexity only when needed. Agents usually trade speed and cost for better results on hard tasks. If a single well-written prompt does the job, you do not need an agent.
The risk side
The more an agent can do, the more it can do wrong. An agent with access to your email can send one. An agent that reads web pages can be tricked by hidden instructions on those pages, a problem called prompt injection. Prefer agents that ask before irreversible actions; Mistral says Vibe Work asks for confirmation before sensitive actions such as sending email, posting messages or changing data in external tools. Read how to use AI at work safely before connecting work accounts.
Try it: a safe first agent task
Pick a task that only reads public information and produces a file you will check. Then follow these steps.
- Open the agent mode your plan includes, for example ChatGPT Work, a Claude task or Perplexity Computer.
- Do not connect email or calendar yet. Keep the first run to public web pages.
- Paste the prompt below, filling in the brackets.
- Watch the steps it shows you. Note where it searched, which pages it opened and where it hesitated.
- Open three of its sources yourself and check the figures against the spreadsheet.
This gives you a feel for how far you can trust the agent before you give it anything with consequences.
Common misconceptions about AI agents
- "An agent is just a smarter chatbot." The model may be the same. The difference is the loop and the tools: an agent can take actions, not only write text.
- "Agents work unsupervised, so I can stop checking." Every maker that ships one still asks you to review results. Treat the output as a draft from a fast junior colleague.
- "Every automation is an agent." A fixed sequence such as "when a form arrives, summarise it and post to Slack" is a workflow. Tools like Zapier, Make and n8n are often the better fit for those, because the steps never change.
- "Agents are free to run." Long agent runs use far more of your plan's limit than a chat. Perplexity Computer is paid with credits, and dots sit on the top plans.
- "It will only do what I asked." Agents interpret goals. Be explicit about what it must not do, such as "do not send anything" or "do not delete files".
Related terms
- Large language model (LLM) is the model that plans each step.
- Computer use covers agents that click and type on a screen. Claude offers it as a beta on Pro, Max and Ultra with Claude Desktop open.
- MCP (Model Context Protocol) is the open standard many agents use to connect to tools and data.
- Deep research is a narrow research agent built into most assistants.
- RAG explains how agents and chatbots pull in your own documents.
- Hallucination explains why agent results still need checking.
Prompts to copy
A safe first agent task
Goal: build a spreadsheet comparing the public pricing pages of [competitor 1], [competitor 2] and [competitor 3]. Columns: company, plan name, what the plan includes, link to the page, date you checked it. Use only the companies' own websites. Do not sign in anywhere, do not fill in any forms and do not send anything. If a figure is unclear, write "unclear" rather than guessing. Show me the list of pages you used at the end.
You get: A spreadsheet with a source link for every row, plus a list of pages you can spot-check.
Ask the agent to plan before it acts
Before you start, write out your plan for this task as numbered steps, list every tool or app you will use and mark any step that changes or sends something. Wait for me to approve the plan. Task: [describe task].
You get: A step-by-step plan you can approve or edit, so nothing irreversible happens without you.