Not Using AI Agents Yet? A Five-Minute Guide to Getting Started
Why agents beat chatbots, plus quick setup tips for Claude Cowork, Claude Code, the Codex app and Codex CLI, including costs, caveats and a sample workflow.
This article is published in English.
Sources: Still not using AI Agents? Get Caught up in Five Minutes.
If you've only used chat-style AI so far, this is a nudge to try an agent. Plenty of people hunt for AI shortcuts and quick tips, but nothing replaces actually working with one. Agents have changed how a lot of people work, build and create, and the gap between those who've tried them and those who haven't keeps growing. A few minutes of reading and a short setup get you started.
What an agent gives you that a chatbot doesn't
Next to a regular chat experience like ChatGPT, it's fair to ask what an agent adds. Four things stand out.
Files that persist
An agent writes real files on your computer, which means it can do legitimate work at your direction. Drop some files in a folder, tell the agent how to transform them and it gets going. The output lives on your disk, not in a scrolling chat window.
Browser automation
You can connect an agent to a real Chrome session. From there it can look at your LinkedIn profile, send messages on X and take care of tedious chores in other software.
Tools, including ones it builds itself
Agents automate complex jobs with built-in tools, and they can write new tools on the spot. Say you check your inbox every day for one particular type of message. Ask the agent to build something for that, and it'll walk you through setting it up. The chore is now handled by an agent that can adjust its own process whenever you ask. It's one prompt away.
Repeatable workflows without coding
Agents usually support skills. A skill is a plain-language description of a task or group of tasks, stored so it can be reused. You, or the agent, write it once and any agent can later pick it up and do the work without a fresh prompt. Suppose you want an agent that audits spreadsheets. Put the exact process into a skill.md file (the agent will happily help draft it) and from then on it repeats that audit on demand.
Getting started
Here's a quick route into agents, beginning with a few of the better platforms.
Claude Cowork
Cowork is the friendliest entry point, since it's a desktop app instead of a terminal tool. These habits help:
- Work inside a project. Start each session in a folder, which Cowork calls a project, so your files stay organized. Choose the option to work in a project, create a folder and you're set.
- Keep critical files out of reach. Never ask it to open or edit your important originals directly. Use copies, or have Cowork copy them into the project. A setting that stops edits outside the project folder is on by default.
- Open up internet access. In settings, set the domain allowlist to all domains. With the restriction left in place, what you can do is quite limited.
- Ask what it can and can't do. Cowork usually knows when a task is out of reach and says so, as long as you ask.
- Request tutorials. When it needs access to your data through a connector or API, ask what it requires from you and it'll arrange it. Be aware that some obvious capabilities are missing from the default connectors. For instance, the Gmail connector in Cowork can't attach files to emails. That's where Claude Code comes in.
- Pick the strongest model. The advice here is to use Opus 4.7, Anthropic's frontier model at the time of writing, for everything. The smaller models, Sonnet and Haiku, are harder to work with, hallucinate more and make more mistakes.
- Add the Claude for Chrome extension to automate small browser tasks. It can burn through credits quickly, so save it for the jobs that matter most.
A sample workflow
A good demo is a repeatable PDF report generator. Asked for one, Cowork starts by interviewing you about the use case. In this example it searches the web for live market data and turns it into a report. You could just as easily wire in your own data source over an API, or through the default connectors that reach Gmail, Google Drive, Notion and similar services.
From there it works on its own. It installs a Python package, searches the internet and starts building the pipeline. After a few minutes it has written scripts and set up a workflow it can rerun every day with no further prompting, one that it or another agent can invoke at any time.
Everything persists. You can ask for changes to the design and it makes them, then repeats the process the next day if you ask. With the Chrome extension enabled, you can go further and tell it to post the finished image to LinkedIn along with a title and body it writes.
Claude Code
Claude Code is nearly the same thing as Cowork, with a different interface. It's a command-line tool, so you talk to it by typing in a terminal.
Out of the box it's considerably more powerful. It can get around many of the guardrails Cowork has, which opens up building harnesses around itself, creating custom software and deploying it to the cloud, and running automated deep web research.
Cowork has Claude Code built in, which is handy if you'd rather not touch code yourself. Technical users will get more from installing it globally and running it in an IDE, and its CLI design has some nice features of its own. Getting going takes a single command in a terminal on Windows PowerShell:
irm https://claude.ai/install.ps1 | iex
That handles installation, setup and sign-in. As with Cowork, it tells you what it can and can't do, and project folders are still the best way to organize work.
Caveats
Claude Code can be a daily driver, but it has real downsides:
- Cost. Expect to spend at least $50 a month for serious use.
- Risk to your machine. Run carelessly, it can permanently damage your computer, overwriting or deleting important files. Read up on running it safely before you give it free rein.
- Limits. It can build a tool for nearly anything, but it can't do everything by itself. You'll find the boundaries as you use it.
With great power comes great responsibility. People have used it for automated trading algorithms for prediction markets, personalized news alerts, small free software projects and simple full-stack applications for themselves and clients.
OpenAI Codex app
The Codex app resembles Cowork most closely, so it's the natural pick if you're already settled in the OpenAI ecosystem. Many people consider OpenAI's current top model, GPT-5.5, the most intelligent available.
Instead of repeating the same demo, a better showcase is its automations feature. A single line of instruction sets up a daily news alert that uses web search. For heavier workflows with complex file reads and writes or data pulls, ask Codex to interview you. Telling it to walk you through using it for a given task is enough to get started.
Codex CLI
OpenAI also has a coding-focused offering, the Codex command-line tool, which has attracted a lot of attention and passed 100 million downloads. It's essentially the same idea as Claude Code, so the choice comes down to preference, and the same principles apply.