AI agents
AI agent examples you can actually use this week
Six small agents for one person: inbox triage, a follow-up reminder, a price or stock watch, meeting notes to tasks, a weekly review and a research digest, each with a copyable instruction.
- The most useful AI agents for one person are small: they watch one thing, do one thing, and have one rule they must never break.
- Six examples below, each with what it watches, what it does, what it must never do, and a plain-language instruction you can paste into whatever tool you use.
- Two of the six need no AI at all; built-in features of Gmail, Google and your phone cover them.
- The rest need a small automation tool (Zapier, n8n) or a coding agent like Claude Code. We say which, honestly.
If you have not met the word before, read what an AI agent is first. The short version: an agent takes a goal, picks its own steps, uses tools, and stops. Anthropic's guide to building effective agents adds the advice that matters most here: find the simplest solution possible and only add complexity when needed. So every example starts with the simplest tool that does the job, and says plainly when a real agent is not required.
Each example has four parts: what it watches, what it does, what it must never do, and a copy block with the instruction in plain language. The instruction is written so you can paste it into a Zapier or n8n AI step, a Claude Code session, or a note to a person. Tooling needed is stated at the end of each one.
1. Inbox triage
Watches: every new email.
Does: labels it as person, receipt, newsletter or notification, archives the last three, and leaves only people in the inbox. Once a day it lists what it archived.
Never: deletes, replies, or unsubscribes.
For each new email, decide which one it is: a message from a real person who expects a reply; a receipt or invoice; a newsletter; an automatic notification. Label it with that word. Archive everything except messages from real people. Never delete, never reply, never click unsubscribe. At 18:00 send me one list: sender and subject of everything you archived today.
Tooling, honestly: most of this needs no AI. Gmail's filter rules can label and archive by sender or keyword. The one part that needs judgement is "is this a real person expecting a reply", and for that you need an AI step in Zapier or n8n, or a coding agent with mail access. Start with filters; add the AI step only for what the filters cannot sort.
2. Follow-up reminder
Watches: messages you sent that end in a question or a request.
Does: if no reply arrives within three working days, puts the thread back in front of you with a one-line summary of what you asked.
Never: sends the follow-up itself.
Look at messages I sent in the last three working days that ask the other person for something. If a thread has no reply after three working days, show it to me again with one line: who, what I asked, when. Do not send anything. If I say "drop it", stop watching that thread.
Tooling, honestly: Gmail already does a version of this without any AI: snooze brings a thread back to the top of your inbox at a time you pick, so snoozing your own sent question for three days is the whole agent. The AI version only earns its place when you send many requests a day and forget to snooze.
3. Price or stock watch
Watches: one product page, once a day.
Does: compares the price to your limit and the availability to "in stock"; tells you only when something changed in your favour.
Never: buys, adds to a basket, or enters payment details.
Every morning at 08:00, open this product page: [paste URL]. Read the current price and whether it is in stock. If the price is below [your limit] or the item has come back in stock, send me one message with the price, the stock status and the link. Otherwise send nothing. Never add the item to a basket and never enter any payment or address details.
Tooling, honestly: for news and mentions rather than prices, a Google Alert sends an email when new results appear, no agent required. For a specific product page you need something that can fetch the page on a schedule: an n8n or Zapier workflow with an HTTP step, or a Claude Code scheduled task on your own computer (more on that in example 6). Many shops also offer their own back-in-stock email, which is the simplest tool of all.
4. Meeting notes to tasks
Watches: a notes file or folder where you drop notes after each call.
Does: pulls out every "who does what by when", writes them as tasks, and shows you the list to confirm before anything is added.
Never: adds a task without your confirmation, or assigns a task to another person.
Read the new notes file. List every action item as: owner, action, due date if stated, otherwise "no date". Only include items the notes say someone agreed to do. Show me the list and wait. When I say "add", put the items with owner "me" into my task list. Never add items for other people and never send the list to anyone.
Tooling, honestly: this is a good fit for a coding agent, because it reads files and writes to a task list through tools. Claude Code works in any folder, including a notes folder, and can be connected to task tools. A Zapier or n8n flow with an AI step also works if your notes land in an app they can read. Amili, the assistant these pages belong to, sorts what you tell it into to do, to remember, or just a thought; it is in a private, invitation-only beta from 14 October 2026, free during the beta, and does not do deep writing or coding. Details on how it works.
5. Weekly review
Watches: your task list and calendar.
Does: on Sunday evening, writes a short summary: finished this week, slipped, due next week, and one question for you.
Never: reschedules or closes anything.
Every Sunday at 19:00, read my task list and next week's calendar. Write at most ten lines: three things I finished, anything that slipped past its date, what is due next week in date order, and one question you would ask me if you were my assistant. Change nothing. Do not move, close or create tasks.
Tooling, honestly: this needs an AI step plus access to your tasks and calendar. If your task app has a Zapier or n8n connector, a weekly scheduled workflow with an AI summarise step is enough. On a Mac or Windows machine, Claude Code's desktop app can run a scheduled task on a weekly schedule, but note the limit in the docs: local scheduled tasks only fire while the app is open and the computer is awake.
6. Research digest
Watches: three to five sources you name (a few sites, a feed, a search term).
Does: once a week, reads what is new, keeps only items that match a question you wrote down, and sends a digest with a one-line reason per item and the source link.
Never: quotes more than a sentence from any source, or presents a summary as fact without the link.
My question: [write the question you are researching]. Every Friday at 16:00, check these sources for anything new since last Friday: [list]. Keep only items that help answer my question. For each one give the title, the link, and one line on why it matters. No more than eight items. Never copy more than one sentence from a source. If nothing new matched, say so in one line.
Tooling, honestly: a Google Alert covers the "search term" source for free. For the rest you want a tool that can fetch pages and summarise: n8n (self-hosted or cloud), Zapier with an AI step, or Claude Code on a schedule. Claude Code offers both local scheduled tasks and cloud routines; routines run on Anthropic's infrastructure so they keep working when your laptop is closed, and the docs list them as available on paid plans as of October 2026.
Which tool should I start with?
| If you | Start with | Needs AI? |
|---|---|---|
| Only want mail and alerts handled | Gmail filters, snooze, Google Alerts | No |
| Want a few apps talking to each other | Zapier (hosted) or n8n (self-hosted or cloud) | Only for the judgement steps |
| Are comfortable in a terminal and want full control | Claude Code with a scheduled task or routine | Yes |
| Want to type or speak and have it sorted for you | An assistant such as Amili (beta, invitation only) | Yes |
Whatever you pick, keep the first agent to one job, run it by hand a few times, and read what it did before you let it run unattended. The "never" line in each example is not decoration; it is the rule that makes an agent safe to leave alone.
Questions people ask
How do I build an AI agent without coding?
Pick one of the two "no AI" examples first (inbox triage with filters, or the follow-up with snooze) so you learn the shape: watch, do, never. Then move one example into Zapier or n8n, where a trigger starts the workflow and an AI step makes the single judgement call. You write the instruction in plain language, exactly like the copy blocks on this page.
Are AI agents safe to run on my email?
They are as safe as the permissions you give them. Give read and label access, not send or delete, and the worst case is a mislabelled email. Every example here has a "never" rule for a reason: an agent that cannot send, pay or delete cannot do lasting damage.
What is the simplest AI agent for beginners?
The research digest or the weekly review, because both only read and summarise. Nothing they produce changes anything, so a mistake costs you a bad paragraph, not a sent email. Run one of them for a month before you give any agent a tool that acts.
Do I need Claude Code or ChatGPT to run these?
No. Two of the six need no AI at all, and the others run in any tool that can schedule a trigger and call a model, including Zapier and n8n. Claude Code is the right choice when you want the agent to work with files on your own computer or you want to read and edit exactly what it does; see a Claude Code workflow for a one-person business.
About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.