First Agent
Build your first AI agent
One question tells you whether you have an agent or a chatbot. After that it is one job, one connection, and one boundary. You do not need a team of them, and the people who build these for a living will tell you the same thing.
Sources opened and quoted on 17 September 2026
The question that settles it
Ask yourself one thing about the AI tool you are using right now.
Did anything change outside of the chat?
Chatbot
You ask it to help you draft an email. You bat it back and forth. Then you copy the output and paste it somewhere yourself.
Nothing moved. You were the hands.
Agent
You ask it to go into your Gmail and draft a response to something you received. You ask it to make the file. You hand it the goal and it does the steps.
Something changed outside the chat window.
That is the whole distinction. A chatbot talks. An agent takes a goal and does the steps. Here is how I say it when someone asks me on a call:
“It’s basically, like, software that can take a goal from you, and it can figure out how to get to the desired outcome by looping through it and actually analyzing its next step without you having to tell it every single thing. That’s all an agent is.”
Anthropic’s engineers draw the same line, and they draw it against workflows rather than chatbots:
“Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
You do not need a team of them
You have seen the diagrams. An AI chief of staff, twelve AI employees, each with a name and a job title. Almost none of it is doing anything.
You do not have to take my word for it. This is Anthropic, writing for the engineers who build these systems professionally:
“we recommend finding the simplest solution possible, and only increasing complexity when needed.”
“Start with simple prompts, optimize them with comprehensive evaluation, and add multi-step agentic systems only when simpler solutions fall short.”
So build one. One job, done properly, that you actually use. That is further than most people with an org chart full of robots.
Pick the job before you pick the tool
This is the step people skip, and it is the step that decides whether the thing is useful. Run your idea past these four. If it fails one, pick a different job.
You already do it by hand
Do not start with something you have never done. You cannot tell whether the output is right if you have no idea what right looks like.
It comes round again
Weekly is a good floor. A job you do once a year is not worth setting up, and you will have forgotten how it works by the time it comes back.
You can describe a finished good one
If you cannot say what a good result looks like in a sentence, the agent cannot check its own work. This is the step most people skip.
Being wrong is cheap
Your first one should not be the thing that goes to a client, moves money, or cannot be undone. Pick something where a bad draft costs you two minutes.
Do not start from a blank page
You know your job better than any AI does. You have just never written it down, so the setup stalls at the first empty box.
So do not write it. Get interviewed instead. Paste this into ChatGPT, Claude, or whatever you use. It asks you six questions and hands you the instructions at the end.
The interview prompt
I want to set up my first AI agent. Interview me before you build anything. Ask me one question at a time and wait for my answer. Work through these six, in order: 1. The job. What is one task I do by hand most weeks? Make me walk you through one real example from start to finish. 2. The trigger. What tells me it is time to do it? 3. The inputs. What do I have to look at to do it, and where does each one actually live? 4. The output. What exists at the end? A draft, a file, a message, a row in a sheet? 5. The finish line. How do I know it was done well? Ask me for one good example and one bad one. 6. The boundary. Which steps may you do on your own, and which have to come back to me before anything leaves? Rules: - One question at a time. Keep the questions short. - If my answer is vague, ask me for a concrete example before you move on. - Do not suggest any tools until you have asked all six. - Do not build anything until I say go. Then give me back three things: - The job written as instructions I can paste into a project or a custom instructions box. - A list of exactly which connections or permissions it needs, and what it can still do for me if I grant none of them. - The one step you recommend I keep as a draft rather than let you send. Use my words. Do not tidy my voice into something more corporate.
Answer it out loud if that is easier, then paste your answers in. What comes back is your first version. It will be wrong in places. You fix it the second time you catch yourself correcting the same thing twice.
Turn on one connection, not nine
A connection is what lets the AI reach something outside the chat. It is the part that turns the answer into an action. Anthropic describes it plainly:
“Connectors let Claude access your apps and services, retrieve your data, and take actions within connected services.”
Grant one. The one your chosen job needs. You can always add another next week.
In Claude
These are Anthropic’s own steps, in their own words:
- 1Click the "+" button in the lower left corner of your chat, or type "/" to open the menu.
- 2Hover over "Connectors."
- 3Select "Manage connectors."
- 4Click the "+" button next to Connectors.
- 5Click the connector you want, then Connect or Install, and follow the sign-in prompts.
Gmail, Google Calendar, and Google Drive are “available for all users on Claude and Claude Desktop”. On a Team or Enterprise plan an owner has to switch them on for the organisation first. Anthropic also states: “We do not train our models on your Gmail, Drive, or Calendar connector data”. Google Workspace connectors.
In ChatGPT
I am not going to give you steps I could not check. OpenAI’s help pages block automated reading, so I could not open them to quote the current wording, and a half-remembered menu name sends you round in circles. Open the app and look in two places: the tools menu inside the message box, and your settings. The six questions above, the prompt, and the boundary rule below are the same either way. Only the clicking differs.
Drafting and sending are two different jobs
This is the rule I run my own business on, and it is the one thing I would ask you not to skip. Write down what the agent may do on its own, and what has to come back to you before it leaves the building.
Reading is low risk. Drafting is low risk. Sending, posting, paying, and agreeing are not, because none of them can be taken back.
- Read my inbox and tell me what needs an answer
- Write the reply and leave it as a draft
- Make the file, the summary, or the list
- Send the reply without showing me. Bring it back to me first
- Message anyone I have not already agreed to message. Bring it back to me first
- Spend money, or agree to anything on my behalf. Bring it back to me first
The good news is that the default is already on your side. On the Gmail connector, Claude can “Send, reply to, and forward emails from Gmail”, and Anthropic says that “By default, Claude asks for your approval before each of these actions.” Leave that alone. It is not friction, it is the thing standing between a helpful draft and an email you cannot unsend.
What I could not check
- I could not open OpenAI's help pages to quote them. help.openai.com and openai.com both refuse automated reading, and they answered 403 every time on 17 September 2026. So this page carries no quoted claim about what ChatGPT can do. The six questions, the prompt, and the boundary rule work in either tool. For where to click in ChatGPT, open the app and look under the tools menu in the message box and under your settings, because that is the only source I can honestly point you at.
- Anthropic's connector pages describe what the feature is meant to do. They are not a test of whether it works well on your inbox, and I have not measured that.
- The Gmail approval step is described as the default. Defaults change, and on Team and Enterprise plans an administrator can change it. Check what yours is set to rather than assuming.
- "Building effective agents" is written for engineers building software. I am borrowing its definition and its warning about complexity, both of which stand on their own. The rest of that post is about code you are not writing.
The agent is the easy part
Setting one up takes an afternoon. Getting a good one takes something else, and it is not a better model.
An agent with no context is a new hire on day one with no briefing. It does not know what you sell, who you sell it to, how you sound, or what it is not allowed to touch. You can hand it all of that once, in plain files you own, instead of explaining yourself in every new chat.
Build your folder brain is the next thing to read. Four questions, four files, and every agent you make after this one stands on top of them.