Agent Zero · main guide

Ten jobs for your agent

These are the ten longer jobs promised at the end of Step 3, and the agent is still Agent Zero, working inside its container. The jobs match the notes from 14 November 2026, and on Windows and on a Mac you paste exactly the same words.

Each job is a goal rather than a single instruction, so paste it, fill in the blanks, and let the agent take more than one step to finish it.

When a job stalls or does something you did not expect, use When something goes wrong.

How to ask

A good request says three things, in this order, and every job below already does all three.

  1. The finished thing. Name the file and say what goes in it, such as a spreadsheet with three rows, a one-page brief with the open questions, or a checklist you can open again.
  2. Where to save it. Say "in the workspace," so the file lands inside the container in the same place as the files from Step 2.
  3. What not to touch. These are the safety rules for every job:
    • Stay in the workspace and the browser.
    • Leave Host access off. If the agent asks to turn it on, or to use files on your real computer, click Stop.
    • Do not send email.
    • Do not buy or book anything.
    • Stay logged out of email, the bank, and stores.

Each job below ends with one short line that repeats these rules, so you do not have to remember them.

For more on asking well, including how to change your requests for a weaker or local model, see Prompting strong and weak models.

The rules are there because of prompt injection, which can hide in any page the agent opens. When you stay logged out, hidden instructions have no account to open, and when Host access stays off, the work stays inside the container no matter what a page says. The words you type are what should steer the agent.

1. Compare three products and save a sheet

Words you can paste

Compare [first product], [second product], and [third product]. Save a spreadsheet in the workspace. Give each product its own row. Include the price, one strength, one drawback, and the web address of the page each fact came from. Use public pages in the browser. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

What the agent should do

  1. Open a public page for each product.
  2. Take the price, one strength, and one drawback from those pages, and keep each page address.
  3. Save the spreadsheet in the workspace.

What finished looks like

The chat says the spreadsheet is in the workspace, and you can open it with LibreOffice Calc after you click the Desktop icon, the same way as in Step 2. Each row has a price, a strength, a drawback, and a page address. Your Documents folder is unchanged, and nothing was added to a cart.

2. Turn a messy expense list into a tidy table

Words you can paste

Turn the list below into a tidy table in a spreadsheet. Save it in the workspace. Give each expense a row with the item and the amount. Read amounts that are written as words. Add a total at the bottom. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

coffee $4.50 LUNCH 12 bus ticket 3.50 yesterday groceries - twenty dollars on Saturday

You can swap in your own list, but keep the sentences above it so the agent still gets the instructions and the safety rules.

What the agent should do

  1. Read every line, including amounts written as words.
  2. Put each expense in its own row, then add the amounts into one total.
  3. Save the spreadsheet in the workspace.

What finished looks like

The chat says the spreadsheet is in the workspace, and you can open it with LibreOffice Calc from the Desktop icon. For the sample list, the four amounts add to $40. For your own list, the total should match what you get when you add the amounts yourself.

3. Plan a weekend trip and save the sources

Words you can paste

Plan a weekend trip starting from [your city]. Save it as a document in the workspace. Include how to get there, a place to stay, and two things to do. For each one, include the price and the web address of the real page you used. Use the browser. Stop before any booking. Follow the safety rules: stay in the workspace, do not send email, and do not sign in or buy anything.

What the agent should do

  1. Open real pages for travel, a place to stay, and two things to do.
  2. Write a simple weekend plan from those pages, keeping the price and the page address next to each item.
  3. Save the document in the workspace.

What finished looks like

The chat says the document is in the workspace, and you can open it with LibreOffice Writer from the Desktop icon. Every price and plan item has a page address you can open yourself to check it. Nothing is booked, and no email was sent.

4. Turn a long article into a one-page brief

Words you can paste

Read the article below. Save a document in the workspace that is about one page. Start with the point of the article, then the few facts that matter, then the questions the article leaves open. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

[paste the whole article here]

What the agent should do

  1. Read the whole article you pasted.
  2. Write a short brief a person can finish in a few minutes, and add the questions the article does not answer.
  3. Save that document in the workspace.

What finished looks like

The workspace holds a short document with the point, the facts that matter, and the open questions, and you can open it with LibreOffice Writer from the Desktop icon. The chat can show the same brief if you would rather read it there.

5. Watch a price, and leave a note

Words you can paste

Watch this page once a day: [page address]. Each time, save a note in the workspace with the date, the price or the main change, and the page address. Use the browser. Do not email anyone, and do not send a text. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

The computer has to stay on, and Agent Zero has to keep running, or the check cannot run.

What the agent should do

  1. Create a daily task.
  2. Each time the task runs, open that page and save a short note in the workspace.

What finished looks like

The check is listed under Tasks, the same area as the morning digest in Step 3. After each run, a note in the workspace gives the date, the price or the change, and the page address. If the computer is shut down, that run does not happen, but either way nothing was purchased and nothing was emailed.

6. Remember how you like summaries

Send these as two separate messages, because the second one is the proof that the first one worked. This is a new preference, saved beside the one from Step 2.

Words you can paste

First message:

Remember this preference. When I ask for a summary, use four short lines, put the most useful fact first, and skip any greeting. Save that preference.

Next message, after the agent says it has saved the preference:

Summarize the note below. Use the summary preference you saved.

The community garden opens at 8 on Saturday. Plots are free for neighbors who live within ten blocks. Tools are in the shed, and the hose was broken last week.

What the agent should do

  1. Save the summary preference.
  2. On the next summary, follow that preference without being told the rule again.

What finished looks like

The second reply is four short lines with the useful fact first and no greeting. A search in the memory dashboard from Step 2 finds this preference, although the notes do not say which of the four areas it uses.

7. Build a checklist you can reuse

Words you can paste

Build a packing checklist for a weekend trip to [place]. Save it as a file in the workspace that I can reuse. Group clothes, toiletries, and papers. Put a box I can check beside each line. Then open the saved file and tell me the groups that are in it. Follow the safety rules: stay in the workspace and do not sign in or buy anything.

What the agent should do

  1. Write the checklist in those three groups for the place you named.
  2. Save the file in the workspace, then open the saved file and report the groups back.

What finished looks like

The chat names the file in the workspace, and the groups match what you asked for. Later you can paste: "Open the packing checklist in the workspace and reuse it for [next trip]."

8. Research a purchase, and stop before paying

The agent must not buy anything, so this job ends at the point where a person would reach for a card.

Words you can paste

Research a purchase of [item]. Save a note in the workspace with two or three options. For each option include the price, the seller, and the page address. Use public pages in the browser. Stop before paying. You must not buy anything. Do not sign in, do not add it to a cart, and do not type a card number. Stay in the workspace.

What the agent should do

  1. Open public pages for the item.
  2. Write down two or three options with the price, the seller, and the page address, and save them as a note in the workspace.
  3. Stop before any payment, sign-in, or cart.

What finished looks like

The note is in the workspace, the browser is not on a payment page, nothing is in a cart, and no card number was typed.

9. Split a comparison across subordinate agents

Agent Zero can hand parts of a job to subordinate agents. Each one works in its own chat with its own instructions, then reports back to the main agent, which writes the final file. The main agent decides for itself when to hand work off, so this prompt asks for the subordinates by name. It also names the researcher profile, one of the agent profiles that comes with Agent Zero. A subordinate can be given any profile that is available in your chat.

Whether the three work at the same time depends on how the main agent calls them. A plain call makes the main agent wait until that subordinate reports back, so three plain calls run one after another. Agent Zero also has a tool named parallel that starts several subordinates together so they work at the same time. The prompt asks for it, but the model still makes the choice. The file comes out the same either way, and the one-after-another route just takes longer.

Before you start: set a spend limit at your provider. Every subordinate holds its own conversation with the model, so this job can use many times the tokens of a normal job.

Words you can paste

Compare [first option], [second option], and [third option] for [what you need it for]. Start three new subordinate agents with the researcher profile, a fresh one for each option, and start them together with your parallel tool if you can. Tell each subordinate to use public pages in the browser and to follow the same safety rules. Each one should report the price, two strengths, two drawbacks, and the web address of each page it used. When all three have reported back, write one comparison document in the workspace yourself. Give each option its own section, then add a table that puts the three side by side and a short paragraph on which option fits my need best and why. Follow the safety rules: stay in the workspace, do not send email, and do not sign in or buy anything.

The line about the safety rules matters, because a subordinate sees only the message the main agent writes for it, not your whole chat.

What the agent should do

  1. Start three new subordinate agents with the researcher profile, one for each option, and give each one its task and the safety rules.
  2. Wait for all three reports. With parallel they work at the same time, and with plain calls they finish one after another.
  3. Write the comparison from the three reports, keep each page address, and save one document in the workspace.

What finished looks like

The chat shows a Calling Subordinate Agent step for each option, and the chat says the document is in the workspace. You can open it with LibreOffice Writer from the Desktop icon. It has a section for each option, a side-by-side table, and a page address beside each price, strength, and drawback. In the chat list, the subordinates' chats sit under the main chat, so you can open one to see what it read. If only the main agent did the work, the file may still be fine, but you can paste the words again in a new chat. Nothing was bought, and you are still logged out.

10. Write a deep-research report

Agent Zero has no button or mode called deep research, so this job asks for one in plain words. That means many sources, a source for every claim, and a clear note wherever the sources disagree. The researcher profile already tells the model to trace each claim to its original evidence and to cite it next to the text it supports. It also tells the model to write down conflicts it cannot settle. That profile writes reports as web pages unless it is told otherwise, so the prompt asks for a document instead.

Before you start: set a spend limit at your provider. A deep-research job reads many pages, and every page it reads goes into the prompt, so this job can use many times the tokens of a normal job. More pages also mean more chances of prompt injection, which is why the safety rules stay in the prompt.

Words you can paste

Research this question: [your question]. Hand the research to a subordinate agent with the researcher profile, and tell it to follow the same safety rules. Use at least eight public sources, and prefer the original source over pages that only repeat it. Save the report as one document in the workspace, not as a web page. Start with a summary of no more than one page, and put the web address of a source after every claim. Then list each point where the sources disagree, and say which sources say what. End with the open questions the sources do not answer, and a numbered list of every source with its web address. If you cannot find a source for a claim, leave the claim out. Follow the safety rules: stay in the workspace, do not send email, and do not sign in or buy anything.

A wide question can also be split, the way job 9 splits a comparison, although each extra subordinate costs more tokens again. To do that, add this sentence just before the line with the safety rules:

Split the question into two or three parts, and give each part its own subordinate agent.

What the agent should do

  1. Start a subordinate with the researcher profile and give it the question and the safety rules.
  2. Let the subordinate open many public pages, keep each claim next to the page that supports it, and report back.
  3. Save one document in the workspace with the summary, the disagreements, the open questions, and the list of sources.

What finished looks like

The document is in the workspace and opens with LibreOffice Writer from the Desktop icon. Every claim in the summary has a page address after it. When you open two or three of those pages yourself, they say what the report says they say. The report has a section on disagreements, even if that section says the sources agreed, and it ends with open questions and the source list. Treat the report as a well-sourced first draft rather than a final answer, because a model can still misread a page. Nothing was bought, and no email was sent.

Sources

Checked 14 November 2026. Source-code links point to the v2.15 release, which was also the newest commit on main that day.

Checked 14 November 2026 · prices, limits, and versions change