How to have an AI agent watch the web for you every morning

An AI agent can check the web for you every morning and send you only what changed: a competitor's new price, a new tender in your field, a mention of your product. You give it one standing search: what to watch, where to look, and what counts as new. With HostNest, a team of Claude agents, you set it up with one message in Telegram, and it runs on your own server, even with your laptop closed.

This guide is for anyone with more to keep an eye on than time to look.

What can an agent watch for you?

Anything you'd otherwise check by hand, again and again. Jobs owners hand over:

1. Competitors. New posts, new prices, new products. "Every Friday, summarise what my three competitors posted this week."
2. Openings. Tenders, grants, job posts, calls for speakers. One builder runs an "opportunity/bid finder for federal contracting".
3. Your market. News about your field, a supplier, or a new law, with a short line on why it matters to you.
4. Mentions. Where your business or your name comes up.
5. What's on. Events in your city that match what you like.

Pick one to start. One clear search beats five vague ones.

Why do agents come back with a mess?

Because the search was vague, or the agent couldn't read the pages. One person asked several agents what's on in their city. His verdict: "They'll run into a wall here, a 403 there, can't parse info elsewhere, javascript somewhere else, and inevitably come back with a weird mix of whatever they could find".

Three things cause most of it:

1. No sources named. The agent searches everywhere and finds noise.
2. No rule for what counts. It can't tell a real change from a page that looks new.
3. No memory of last time. It sends the same things again.

Each one has a fix, in the next section.

How do you brief the search so the results are good?

Write the brief once, and keep it. A good one has six parts:

1. What to watch. "Prices and new products on these three competitors' sites."
2. Where to look. A list of sites. Claude can keep a search to the sites you name: Anthropic's docs say it can "scope results with 0 to include only certain hosts".
3. What counts. "A price change, a new product page, or a new post. Not a changed footer or a new cookie banner."
4. What to skip. "Nothing older than a week. No job ads."
5. Proof for each result. Ask for the link and the exact line that changed. A summary can miss things: Anthropic's docs warn that reading a page this way is "lossy by design", so the quote is your check.
6. Only what's new. "Keep a list of what you've already sent me, and skip anything on it."

Here is one you can copy and change:

Every weekday at 7:00, check these sites: [your three competitors].
Send me only what changed since yesterday: a new price, a new product, or a new post.
For each one: the link, the exact line, and one sentence on why it matters to me.
Skip anything you've already sent me. Keep a list of what you've sent.
If a site won't load, say so. Don't guess.
If nothing changed, send one line: "Nothing new today."

How do you make it run every morning?

It has to run as a job on a computer that stays on. A chat window can't do it: when you close it, it stops.

With HostNest, you send that brief to your agent in Telegram. Your agent sets it up as a job on its own server, and the job survives restarts and updates. Each morning the results arrive in your chat, from that agent. If you'd rather it ran weekly, say "every Friday". Our guide on giving an agent a weekly job covers the schedule in more detail.

If you run your own agent, it needs a server that stays on and a scheduler on it, such as cron, that starts the search at the time you set. You also keep the agent and its server updated yourself.

How does it get better each week?

By telling it what was useful, and having it write that down. One person who has an agent search for openings put it simply: "What I want is pretty simple: Hermes searches, I review and rate the results, and Hermes uses that feedback to improve its next searches."

Do it in plain words, right in the chat:

1. Say what was good. "The price change was exactly right."
2. Say what was noise. "Don't send blog posts about hiring."
3. Ask it to update the brief. "Add that to your rules for this search."

On HostNest, your agent keeps its memory as plain files on your server, so a correction you give it today is still there next week.

What can't an agent see?

Pages behind a login, pages that block automated visitors, and paywalled articles. Some sites change often and break a reader. That's why the brief says: if a site won't load, say so, and don't guess. A short "couldn't read site X today" is worth more than a made-up summary.

If a site you need sits behind your login, ask your agent whether it can connect with your own account. Nothing comes connected on day one: an agent connects your tools when you ask, usually with your own login.

How do you get an agent that does this, without setting it up yourself?

With HostNest, a team of Claude agents, you sign up in about 10 minutes and your agent is ready on its own private server in the EU. Then you send it the brief above. Our guide on how to get a personal AI agent walks through the signup step by step.

What's handled for you: updates install automatically between 3 and 5 am server time, and a failed update rolls back on its own. The server is backed up daily. Your agents check with you before anything goes out or costs money, unless you tell them they don't need to. When one search turns into several, you can add a teammate with its own job. Signup starts at hostnest.ai.

What does it cost?

With HostNest, there's a 7-day free trial, then one subscription for the whole team. You also need your own paid Claude plan (Pro, Max, Ultra or Team), paid to Anthropic. A daily search uses some of your Claude plan's limits, so start with one search and see.

If you run your own agent, you pay for the server and the model, and the upkeep is your time.

Which way suits you?

• You want it running tomorrow morning, without a server to look after: HostNest. Sign up, send the brief.
• You like running every part yourself: your own agent on your own server, with a scheduler.
• You only need it once: ask any AI chat to search now, and skip the schedule.

A checklist

1. Pick one thing to watch.
2. Name the sites.
3. Say what counts as new, and what to skip.
4. Ask for the link and the exact line for each result.
5. Ask it to keep a list of what it has sent.
6. Run it as a job on a schedule, not in a chat window.
7. Correct it in plain words, and ask it to add the correction to its rules.