What is an AI agent, actually?
The same kind of AI you chat with, given a goal, tools and permission.
One ask, a dozen steps
You asked for one thing. It took a dozen steps.
The same kind of AI
Given three things:
- A goal
- Tools to use
- Permission to keep going
Then it goes round a loop
Round again, until it's done or stuck.
- It picks a step
- A tool does it
- It reads what came back
First, it picks a tool
Its tools are a short menu. It reads each one, then picks.
Then a tool does the step
The AI only asks. Sending, paying or deleting should wait for your yes.
Then it reads what came back
Thursday's full, so it tries Friday. Nobody planned that step.
Not smarter. Just more steps.
Chat answers once. An agent keeps going. Start small, and read its steps.
My pick for handing over a task
ClaudeAlso good: 
ChatGPT Work
Hand over a task in any chat: it plans the steps, works with your files and asks before important actions.
All my picks
Watch one agent loop
Public information only.
- Paste it, with your town, into Claude, ChatGPT Work or a Gemini task.2 min
- Watch its steps: pick, tool, read, again.5 min
- Spot one step chosen because of a result.3 min
The request
Find three free events for adults at the public library in [your town] in the next two weeks. Use the library's own website. Make a table: date, event, link. Don't sign in, book or send anything. When you finish, list every step you took, in order.
Good looks likeYou can point to each lap of its loop.
In your assistant · as of Oct 2026 · Claude, ChatGPT, GeminiWhere the agent lives
It's in the app you already use. Menus move often.
- CheckAllowed at work
- CheckAsk before acting: on
- CheckA small job first
Never type a password or payment details into a task. When a site needs you to sign in, take over and do it on the site.
Claude
Calls it a task. It works in any chat.
Describe the result you want and attach files. There is nothing to turn on.
Switch on only what the job needs. In the desktop app, choose the folders it may use.
Check its progress, even from your phone
By default, Claude asks before taking an action. Keep it that way.
Since 16 Sep 2026, reaching Pro and Max first, then more plans. Task work uses your limits faster than chat.
ChatGPT
Calls it Work: its agent for longer tasks.
The toggle at the top of the desktop app. On the web or phone, select Work.
Add files, rules and how you'll judge the result.
Answer its questions, change direction, and approve important actions.
On Free and Go, Work is limited, in the desktop app only. The older agent mode is no longer available: use Work instead.
Gemini
Calls it a task: Gemini's agent.
Below the text box. Describe what you want done, and when.
Connected apps, skills and a browser
Your own Chrome on a computer, or a browser in the cloud.
It shows the task's steps. It's built to ask before it sends, buys or submits a form.
A personal Google Account on Google AI Pro or Ultra, 18 or over. Not in the EEA, Nigeria, Switzerland or the UK. At work, start with Gemini in Gmail and Docs.
Quick check · three questionsDid it stick?
1A colleague says their AI “became an agent” this week. What most likely changed?
The answer: C An agent is the same kind of AI. What changes is what it may use, and that it may keep going.2An agent “checks your calendar”. What actually opens the calendar?
The answer: A The AI only asks for the step. The program around it runs the tool and hands back what happened.3Its first choice of day turns out to be full. What happens next?
The answer: B Nobody scripts the steps in advance. It picks each one after seeing how the last one went.
Why long jobs slow downSide note
Agent, or fixed steps?Side note
Who stops the loopSide note
Sources, checked 18 Oct 2026Source
Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.
- Anthropic, Building effective agents (agents are "LLMs using tools based on environmental feedback in a loop"; they direct their own steps; stopping conditions such as a maximum number of iterations; workflows for well-defined tasks; higher costs)
- Anthropic, Effective context engineering for AI agents ("LLMs autonomously using tools in a loop")
- Anthropic, How tool use works ("The model never executes anything on its own"; the model asks, your code runs it, the result flows back; only a tool can perform actions)
- Anthropic, Tool use with Claude (Claude picks a tool from the request and the tool's description)
- Anthropic, Define tools (a plain description of what each tool does and when to use it, 3 to 4 sentences)
- Anthropic, How the agent loop works (the loop; one or two turns for a quick question, dozens of tool calls for a complex task; context accumulates; caps on turns and spend)
- Anthropic, How we built our multi-agent research system (3 to 10 tool calls for simple fact-finding; hundreds of turns)
- Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026; you don't choose where a task goes; any conversation; Pro and Max first, more plans to follow; nothing to turn on; asks before taking an action by default; carries on after you close your laptop; progress from your phone)
- Anthropic, Claude release notes (16 Sep 2026: Claude Cowork comes to every conversation, rolling out on Pro and Max)
- Anthropic, Cowork overview (describe a desired outcome)
- Anthropic, Claude Cowork product page (uses limits faster than chat; you choose the folders and tools; folders on the desktop app)
- Anthropic, Get started with connectors (+ in the message box, then Connectors, per chat)
- OpenAI, A practical guide to building agents (model, tools, instructions; simple chatbots are not agents; irreversible or high-stakes actions need human oversight; start small)
- OpenAI, ChatGPT Work and Codex (Work is an agent for longer, multi-step work; the toggle at the top of the desktop app; select Work on the web or phone; describe the outcome you want; review progress, change direction, approve important actions)
- OpenAI, ChatGPT release notes (9 Jul 2026: Introducing ChatGPT Work, an agent for longer, more involved tasks)
- OpenAI, ChatGPT agent ("ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables.")
- OpenAI, ChatGPT plans (Work: Limited, desktop app, on Free and Go)
- Google, Agent Development Kit, Agents (a model, task instructions and, optionally, tools)
- Google, Manage your tasks and workflows in Gemini Apps (Task below the text box; 18 or over, personal account, Google AI Pro or Ultra; not in the EEA, Nigeria, Switzerland or the UK; local Chrome or a remote browser; work panel; designed to ask before sending, changing data, buying or submitting forms; confirm or take over for sign-ins; no passwords or payment details in the task)
- Google, Get started with Google Workspace with Gemini (Ask Gemini in Gmail and Docs on an eligible work account)
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency (only the tools the job needs; a person approves high-impact actions)
- UK National Cyber Security Centre, Managing the cyber risk of agentic AI (20 Aug 2026; when an agent should stop and seek human approval; do not rely on prompting alone)
Free kitThe Agent PlaybookBring this to your teamAI implementation · Get in touchBack to level 4, Agents
A goal, tools and permission.
Next lesson · 15 minHow do you hand a task to an agent?