Ask a chatbot for invoicing software that suits a freelancer, and it answers with what it already knows. Give the same request to an AI agent, and it opens the vendors' websites, notes the prices, fills in a table, then comes back to tell you what it couldn't confirm. To do that well, it needs a written assignment with a clear finish line.
Chatbot, assistant, agent: what's the difference?
Anthropic, the company behind Claude, defines agents as language models using tools in a loop, based on feedback from their environment. Put simply: the agent takes an action, looks at the result, then decides what to do next.
Claude Code's documentation puts it bluntly: without tools, Claude can only respond with text. With tools, it can search the web, read and create files, and use connected services. "Assistant" is a looser word that gets used for almost anything in between. For fuller definitions, see our glossary entries on agents and agentic AI.
How an agent works, step by step
Claude Code's documentation describes three phases that repeat: gather context, take action, and verify results. The agent decides each step based on what it learned from the previous one. Nobody mapped out the route in advance.
Say a shop owner wants to compare three point-of-sale systems. The agent reads her criteria and looks up the vendors' sites. It notes the prices and fills in the table. Then it goes back to two pages to double-check prices that look off. If it finds a mistake, it goes around again.
You stay part of the loop: the same documentation says you can interrupt at any point to steer it somewhere else. Anthropic uses the word "loop" for this cycle of work that repeats until a stop condition is met, and we cover it in a separate article.
The agents you already have access to
Claude
On September 16, 2026, Anthropic announced that Cowork and chat are now one Claude, rolling out on Pro and Max plans. The mode that works with your files and apps is now part of Claude itself. Our Claude Cowork guide shows how to use it.
ChatGPT
OpenAI's agent mode is called ChatGPT Work. According to the launch post, you can follow its progress, answer questions, change direction, and approve important actions. On web and mobile it opened first to Pro, Enterprise and Edu, then Plus and Business. In the desktop app, it's available on every plan, including Free.
Gemini
Google launched Gemini Agent in November 2025, then Gemini Spark in May 2026, an agent that works on your tasks on your behalf. Both require a Google AI Ultra subscription and started in the United States.
These agents plug into other services (calendar, cloud storage, CRM) through connectors, often using a protocol called MCP, which our MCP guide explains. For a first try, pick a task you already know how to do yourself: you'll spot right away where the agent goes wrong.
Briefing an agent in five lines
You steer a chat one question at a time. An agent goes off and works alone, and anything you leave out, it guesses.
Goal
the objective in one sentence. "Find three meeting venues within an hour of Chicago for 25 people in March."
Context and constraints
what it needs to know and what it must not touch. Budget, dates, sites to avoid.
Done when
how you'll know the job is finished. "Three venues with price, capacity and availability confirmed on the venue's own site." Without this line, the agent stops whenever it decides to. Add a limit too, as Anthropic recommends with a maximum number of iterations: "if you haven't found three after checking 15 sites, stop and tell me why."
Permissions
what it can do alone and what needs your sign-off. Searching and comparing, yes. Booking, paying or sending an email, never without you.
Report
the format of the deliverable and what it should flag at the end.
The human checkpoint goes before the agent starts. Ask for its plan and only let it begin once you've read it. A wrong turn is easy to spot in a ten-line plan and much harder to spot in a finished table.
For multi-step assignments, our advanced prompting course on workflows and agents goes further.
What agents get wrong, and why you stay in charge
An agent runs longer than a chat, and that's when things go sideways.
It clutters its own desk
Every page it reads and every search result it gets is added to its context window. On a long task, the documentation admits that instructions from early in the conversation can get lost. Our article on the context window explains why.
It calls the job done too early
In the simplest kind of loop, the agent itself judges that it has completed the task. So it can hand back a half-checked table with total confidence. In our view, the most useful line in the brief is the last one: "tell me what you're unsure about."
Some actions can't be undone
A draft goes in the bin. A sent email or a payment doesn't. Even in Claude Code, the developer tool, the documentation notes that actions that affect remote systems can't be checkpointed. ChatGPT Work also builds in a step where you approve important actions. Write your own limits into the brief, on top of the vendor's guardrails.
Warning : start with tasks where a mistake costs little, and ask for drafts. For its scheduled tasks, Anthropic advises having Claude draft for your review rather than send on your behalf.
To check an agent's work quickly, see our article on human in the loop.
Frequently Asked Questions
What's the difference between an AI agent and a regular automation?
An automation follows steps someone wrote in advance. An agent picks its steps as it goes, based on what it finds. It copes better with surprises, and it's also less predictable.
Can an agent work while I sleep?
Yes, with scheduled tasks: the assignment starts on its own at the time you set. In Claude, scheduled tasks have run with no device online since July 2026, on paid plans. Our guide to scheduled tasks shows how to set them up.
Where can I find ready-made agents?
Our AI agents catalog collects free, step-by-step workflows you can adapt to your job.
Prompt Guide
Founder of Prompt Guide and CEO of Webpulser. Digital and AI entrepreneur since 2006, he shares his field-tested prompt engineering techniques.
Practice what you learned
Interactive exercises to sharpen your prompting skills
Stay Updated
Get our best articles and techniques every week.
Related Articles
Make AI Finish the Job: The 4 Kinds of AI Loops, Explained Simply
Anthropic calls it a loop when an AI repeats its work until a stop condition is met. Here are the four kinds in plain English, what each looks like in a regular chat, and three prompts you can copy.
Context Window, Memory and Projects: What the AI Actually Sees
An AI only sees what fits on its desk. Context window sizes as of October 28, 2026, memory, projects and overlong chats: how to give it the right context.
Human in the Loop: How to Check AI's Work Without Redoing It
Rereading everything wipes out the time you saved; skipping the check means signing blind. What human in the loop means, how to scale checks to the stakes, and quick ways to verify AI output.
Automate Your Monitoring with ChatGPT, Claude and Gemini Scheduled Tasks
A scheduled task is a prompt that runs itself. What ChatGPT, Claude and Gemini offer as of October 28, 2026, how to set each one up, and three monitoring prompts ready to schedule.
Related Prompts
Developer Resume Prompt
Write a developer resume from your stack, your projects and the target job posting.
From Meeting to Action Plan: Expert Prompt for Consultants
This prompt turns meeting notes into a structured action plan adapted for B2B consulting firms, with levers, budgets, and KPIs.
Summarize a Course to Study Effectively with AI
A prompt that turns any course or set of notes into a structured, clear, easy-to-memorize study sheet, tailored to your level and your deadline.
Write a SKILL.md for Hermes Agent: turn a repetitive task into a skill
Converts a freely described manual task into a structured SKILL.md for Hermes Agent: triggers, preconditions, procedure, checks, failure cases, report.