Back to My First Agent

Start here

Create your first agent by giving it one small job.

An agent is just an assistant that can follow steps, use allowed tools or files, and come back with a result. Start small, keep it safe, and check the work yourself.

The simple idea

A good first agent prepares work. It does not make the final decision.

Small job Choose one repeated task, not a whole department or system.
Safe material Use public, fake, or low-risk material while you learn.
Human review The agent prepares the output. You decide whether to use it.

First agent path

Five steps, no jargon required.

  1. 1

    Pick one job

    Choose a task you already repeat: prepare notes, draft a brief, compare documents, make a checklist, or summarise something.

  2. 2

    Make it tiny

    Shrink the task until it can be tested once with low risk. One document, one meeting, one example, one folder.

  3. 3

    Give clear instructions

    Say what the agent may read, what it should produce, and when it must stop and ask you.

  4. 4

    Ask for evidence

    Tell the agent to show what it used, what it changed, what it is unsure about, and what you must check.

  5. 5

    Review before using

    Do not let the first agent send, approve, delete, pay, update, or publish. Look at the work first.

Example first agent

A meeting-prep agent is a safe place to start.

A simple first agent might read one agenda and one safe notes document, then prepare a short briefing for you to review.

Prepare a meeting brief from these notes. Give me the key points, open questions, risks, and anything I should check. Do not send anything or update any system.

That is enough for a first agent. It has a small job, a limited source, a useful output, and a human review point.

Plain meanings

Use ordinary words first.

Agent

An assistant with a job, instructions, allowed material, and a stopping point.

Tool

Something the agent is allowed to use, such as a file, search, calendar, document, or app.

Instruction

The written steps that tell the agent what to do and what not to do.

Evidence

The notes that let you check where the answer came from and whether it is safe to use.

Final question

What is one repeatable workflow you would trust an assistant to prepare, but not yet decide alone?

Use the wizard