Atomic Agent

Atomic Agent: Nine jobs to try

These are longer jobs that play to what Atomic Agent does well: reading documents, writing files, searching the web, driving a browser, remembering how you work, and running one long job on its own. They come after Phone and schedules, they are ordered from easy to ambitious, and they work the same on Windows and on a Mac.

The jobs match Atomic Agent v0.7.0 and the desktop app 0.0.8, as checked on 14 November 2026. If a job stalls or does something you did not expect, go to the fixes page.

To see why these prompts are worded the way they are, and how to adjust them for a strong cloud model or a small local one, read Prompting strong and weak models.

How to ask

Atomic Agent runs on your real computer, so every job below names one folder and tells the agent to stay in it. Use your atomic-tryout folder, or a new empty folder made for that one job, and paste its path where the prompt says so.

The project's tips page says small models do best with one clear goal, a clear finish line, and a starting point such as a path or a web address. Each prompt below already has all three, plus these rules:

The rules are there because of prompt injection. Web search and page reading never ask for approval, so a hidden instruction on a page can reach the agent without a card, and your answers on the cards that follow are what keep it in check.

1. Sum up a long document

This is the easiest job, because the agent can pull text out of PDF, Word, Excel, PowerPoint, RTF, and ODT files on your computer. Copy one harmless document into the job folder first, such as a product manual or a public report.

Work only in this folder: (paste the folder path here). Read the document in it. Write a note named brief.txt in the same folder that starts with the main point, then lists the five facts that matter most, then lists any open questions. Do not change the original file. Do not touch any other folder.

Done looks like one new file, brief.txt, next to the unchanged original.

Safety note: with a cloud model, the document's text goes to that company, so use a file you would not mind sharing.

2. Turn a messy list into a table

The agent writes plain files well, and a CSV file opens as a spreadsheet in Excel, Numbers, or LibreOffice. The desktop app's side panel can preview a CSV file in the chat as well.

Work only in this folder: (paste the folder path here). Turn the list below into a file named expenses.csv in that folder, with one row per expense and the columns item and amount. Read amounts written as words. Add a last row named total. Do not touch any other folder.

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

For the sample list, the four amounts add to $40, so check that the total row matches what you get when you add them yourself.

Safety note: the only approval this job needs is one file write in your job folder, so deny anything else.

3. Research a question and keep the sources

The agent can search the web and read pages without opening a browser, which makes quick research cheap. Those reads never ask first, so keep the job's output inside the folder and check the sources yourself.

Find out how long a typical laptop battery lasts before it needs replacing. Use web search and read at least three public pages. Write a note named battery-research.txt in this folder: (paste the folder path here) with a short answer, then each source's page address and the one fact you took from it. Do not sign in anywhere. Do not touch any other folder.

Done looks like a note with an answer and three page addresses that you can open to check.

Safety note: a page can carry a prompt injection, so read any approval card that appears during research, and deny anything that is not a write to battery-research.txt.

4. Compare three products in the browser

This job needs Chrome, Edge, or Chromium installed, because the agent drives a real browser window. The browser asks before it submits a form or types into a password or card field, but ordinary clicks and page reading go ahead without asking.

Use the browser to compare [first product], [second product], and [third product] on public store pages. Write a file named compare.csv in this folder: (paste the folder path here) with one row per product and the columns price, one strength, one drawback, and page address. Do not sign in. Do not add anything to a cart. Do not buy anything. Stop if a page asks for a password or a card.

Done looks like compare.csv with three rows, an empty cart, and no sign-in.

Safety note: if a card asks to submit a form or to type into a password or card field, click Deny.

5. Remember how you like summaries

Atomic Agent keeps profile facts in its memory and adds them to later chats on its own. Send these as two separate messages, because the second one is the proof that the first one worked.

Remember this as a profile fact about me: when I ask for a summary, use four short lines, put the most useful fact first, and skip any greeting.

Then, in a new chat:

Summarize this note using my summary preference: 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.

Done looks like four short lines with the useful fact first and no greeting.

Safety note: saving to memory never asks for approval, so never tell the agent a password, key, or account number as a "fact."

6. Sort a copy of your files into folders

This is a bigger file job, so do it only on copies. Copy ten or so harmless files, such as old downloads or photos you have backed up, into a new folder made for this job.

Work only in this folder: (paste the folder path here). Sort the files in it into subfolders named Documents, Pictures, and Other, inside this same folder. Do not delete anything. Do not rename any file. Do not touch any other folder. When you are done, list what went where.

The agent moves files with its own move tool, which makes any missing subfolders and asks once for the whole batch. Read the full list of moves on the card, and click Allow once only if every path in it is inside your job folder. The tool never replaces an existing file unless asked, and its result lists the moves so you can undo them.

An open GitHub issue reports a local model on Windows looping on a file-sorting job like this one, so expect better results from Gemini than from a small local model here.

Safety note: one Allow once on the move card covers every file on it, so the copy in its own folder is what makes a mistake easy to undo.

7. Let one long job run while you are away

An overnight run hands the agent one long job with a time limit and lets it work alone. In the desktop app, open the + menu next to the message box and choose Run overnight, then fill in Task, Folder, and Time limit, hours, which runs from 0.5 to 24.

Pick a new folder made for this job, put a copy of a few notes or documents in it, and set a short time limit, such as 1 hour, for your first run.

Read every note and document in this folder. Write a file named index.txt in the same folder with one line per file: its name, one sentence about what it says, and its date if it shows one. Do not change, move, or delete any other file. Do not use the web. Do not touch any other folder.

In the desktop app, a step that needs your yes waits for you in the app, and a small card in the corner shows the time, with Open and Stop buttons. Keep the app open, because the docs say the app window enforces the time limit.

Safety note: the overnight docs say the desktop dialog's Budget field is not applied yet in this preview, so on a paid model, rely on the spend limit at your provider, not on that field.

8. Write a deep-research report with a source for every claim

Atomic Agent has no separate research mode, but its own site lists research as a use case, where the agent reads pages and PDFs across many sources and writes a cited note. This job is a bigger version of job 3, with more sources, a page address beside every claim, and a list of the places where the sources disagree.

Research this question: [your question, such as whether a heat pump would save money in a cold climate]. Use web search and read at least eight public pages from different publishers. Write a report named research-report.txt in this folder: (paste the folder path here). Start with a short answer. Then list each finding with the page address it came from on the same line. Add a section named Disagreements that lists each point where the sources disagree, with both page addresses, and a section named Gaps for anything you could not find a source for. Do not sign in anywhere. Do not buy or send anything. Do not touch any other folder.

Done looks like a report with a short answer, a page address beside every finding, and a Disagreements section, so open a few of those addresses to check that each page says what the report claims.

The tools page says web search and page reading never ask first and can read several pages at the same time. It also says a loop detector ends the turn gracefully if the agent keeps searching without making progress, which can happen on a small local model.

Safety note: every page the agent reads can carry a prompt injection, so deny any card that is not a write to research-report.txt.

Cost note: if you use a cloud model, set a spend limit at your provider before you start, because reading this many pages can use many times the tokens of a normal job.

9. Split one job between several workers with Fusion

Fusion, from Advanced features, hands one job to an orchestrator that plans and checks the work, and to workers that each do one part. The Fusion page says the workers run "several at once," but local workers run one at a time unless you raise the worker count, and cloud workers are capped at 4 at a time by default.

You need two providers first, usually a cloud provider with a key and a local model that is already downloaded. In the desktop app, open the Where it runs switch in the composer and choose Fusion. Then make a new, empty folder for this job.

Work only in this folder: (paste the folder path here). Compare [first product], [second product], [third product], and [fourth product] using public pages, with one worker for each product. Each worker writes one text file in this folder, named after its product, with the price, one strength, one drawback, and the page address of each fact. Then check the four files and tell me in the chat which product looks best and why. Do not sign in. Do not buy or send anything. Do not touch any other folder.

Atomic Agent asks you once for the whole fan-out, before any worker starts, and the request lists the task titles and the folders the workers may write to and run commands in. Approve it only if the one folder listed is your job folder, because that single answer covers the rest of the turn. While it runs, each worker has its own row with Steer and Stop buttons that act on that worker alone.

Done looks like four files in the job folder and a reply in the chat comparing them. The orchestrator never writes files itself, so the comparison comes back in the chat rather than as a file.

Safety note: approving the fan-out lets every worker write and run commands in the approved folder without asking again, although the README says that approval never covers dangerous commands. Whichever seat runs in the cloud also sees your request and the files it reads.

Cost note: set a spend limit at your cloud provider before you start, because a split job can use many times the tokens of a normal job. The Fusion page documents a money cap for one turn, turnBudgetUsd, but you can set it only by editing config.json, and its costs are estimates from the price list, not your bill.

Sources

Checked 14 November 2026.

Next

Next: Fixes.

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