Hermes Agent

Hermes Agent: Nine jobs for Hermes

This is page 5 of 8 in the Hermes Agent guide. It gives you nine jobs that fit what Hermes's docs say it does well, from easy to ambitious. It assumes the safety switches from page 2 are still on.

Each job is a goal rather than a single instruction. Paste it, fill in the blanks, and let Hermes take more than one step. The jobs match the docs as checked on 14 November 2026. You paste the same words on Windows and on a Mac.

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.

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

How to ask

A good request says three things, and every job below already says all three.

  1. The finished thing. Name the file and say what goes in it.
  2. Where to save it. Say "in the agent-tryout folder," because on the local backend nothing else keeps Hermes there.
  3. What not to touch. These are the limits for every job:
    • Stay in the agent-tryout folder and touch no other folder.
    • Use the file tools only, and do not write or run a program. A Python program's file changes do not ask first, in the desktop app or in a terminal chat.
    • Use web_search and web_extract for web pages, not the browser, because the browser runs a program.
    • Do not sign in, send email, or buy anything.

Keep approvals on manual, read every command before you approve it, and deny anything that reaches outside agent-tryout. The limits matter because of prompt injection, which can hide in any page Hermes opens. On Hermes, a successful trick runs on your real computer.

1. Make an index of your notes

Hermes reads and writes files well, so start with a job that only reads and adds one file. Copy three to five text notes into a new notes folder inside agent-tryout, and keep the originals somewhere else.

Read every text file in the notes folder inside agent-tryout. Write one new file called notes-index.txt in the agent-tryout folder that lists each file name with a one-sentence summary. Do not change, rename, move, or delete any existing file. Use your file tools only, do not write or run a program, and do not touch any other folder.

What you should see notes-index.txt with one line per note, and every note in notes unchanged.

Safety note: the notes are copies, so even a mistake costs you nothing.

2. Answer a question and keep the sources

Hermes's web tools search and read pages without a key. It can answer a factual question and show where each fact came from.

Find out [your question, such as how long a passport renewal takes in my country]. Use web search and read at least two pages. Save a short answer as answer.txt in the agent-tryout folder, and put the web address of the page next to each fact. Do not sign in, do not fill in any form, and do not touch any other folder.

What you should see answer.txt with a short answer and a page address beside each fact. You can open each page yourself to check.

Safety note: treat any page that asks Hermes to sign in or run a command as a reason to deny and stop.

3. Remember how you like summaries

Hermes keeps your preferences in USER.md between chats. Send these as two messages in two chats, because the second chat is the proof that the first one saved anything.

Send this first message.

Use your memory tool to save this preference. When I ask for a summary, use four short lines, put the most useful fact first, and skip any greeting. Do not save anything else.

Type /new, then send the second message.

Summarize notes-index.txt from the agent-tryout folder. Use my saved summary preference, and do not touch any other folder.

What you should see four short lines with no greeting, and a matching line in USER.md.

Safety note: open USER.md now and then and delete anything it saved that you did not ask for.

4. Find something from an earlier chat

Hermes stores every past chat and can search them word for word. It can find a decision you made last week without you scrolling. This works best after a few days of use.

Search our past chats for what we found about [topic, such as USB drives]. Quote the message where we decided, and tell me the date of that chat. Do not run any commands, and do not change any file.

What you should see a quoted message from an earlier chat with its date. There are no approval questions, because a search runs no commands.

Safety note: the search reads only your own chat history. That history can hold anything you once typed, so never paste a password into a chat.

5. Save a routine as a skill

Hermes's signature feature is writing its own skills, which are saved how-tos it reuses. With write_approval: true from page 2, each new skill waits for you, so you can read it before it shapes later jobs.

Make a skill called weekly-notes-index that repeats job 1. It should read the text files in the notes folder inside agent-tryout and write notes-index.txt there. The skill must only read those notes and write that one file, and it must not run programs or touch any other folder.

Then type /skills pending to see the waiting skill and /skills diff to read the whole change. Approve it with /skills approve only if every step stays inside agent-tryout.

What you should see the skill listed under /skills pending before you approve it. Later, a "run the weekly-notes-index skill" request rebuilds the index.

Safety note: reject any skill that names another folder, a login, or a program, because a bad skill repeats its mistake every time.

6. Check a page every day and log it

Hermes runs schedules inside its gateway, so a daily check works only while the computer stays on and Hermes is running. Hermes ships a product-price-monitor skill as a worked example of this kind of job.

Every day at 9am, use web_extract to read [page address], not the browser, and find the price. Add one line to price-log.txt in the agent-tryout folder with the date, the price, and the page address. Do not sign in, buy anything, or send any message, and do not touch any other folder.

What you should see the job listed by /cron list in a terminal chat, or in the desktop app's scheduled jobs view. A new line appears in price-log.txt after each run while the computer is on.

Safety note: leave the scheduled-job setting at its default, which blocks dangerous commands when nobody is there to answer.

7. Get a morning summary on your phone

This is one of the riskiest jobs, because it combines a schedule, the web, and your phone. Do it only after page 4 is done and your allow list holds just your own Telegram account. Hermes's own schedule docs use this example.

Every morning at 9am, check Hacker News for AI news and send me a summary on Telegram. Use public pages only, and keep it to five short lines with a link for each item. Do not sign in anywhere, and do not run any commands or change any files.

What you should see one short summary in your Telegram chat each morning while the computer is on, and nothing in agent-tryout changed.

Safety note: a job that sends you messages runs with nobody watching. Keep it to reading public pages and replying only to you.

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

Hermes's web tools can search and read many pages, so this job is a much bigger version of job 2. Hermes ships a grounded-citations skill for cited reports, but it works by running a Python script, which breaks this page's no-programs limit, so the prompt asks for the same result in plain words.

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 pages from different publishers. Save a report as research-report.txt in the agent-tryout folder. Start with a short answer, then list each finding with the web address of its page next to it. Add a section called Disagreements that lists each point where the sources disagree, with both page addresses, and a section called Gaps for anything you could not find a source for. Use your web and file tools only, do not write or run a program, do not sign in, send, or buy anything, and do not touch any other folder.

What you should see research-report.txt with a short answer, a page address beside every finding, and a Disagreements section. Open a few of the addresses to check that each page says what the report claims.

Safety note: if Hermes asks to run sources.py or any other program, deny it. Every page it reads can carry a prompt injection, so treat a request to sign in or run a command as a reason to stop.

Cost note: 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 a research job across sub-agents

Hermes can hand parts of a job to sub-agents, each with its own fresh conversation and the same tools as your chat. Its delegation docs say a batch runs "up to 10 concurrent subagents by default," so they work at the same time, not one after another. Hermes posts one combined result when every sub-agent has finished.

Compare [first product], [second product], and [third product]. Use sub-agents, one for each product, working at the same time. Each sub-agent should use web search to find the price, one strength, one drawback, and the web address of the page for each fact, and report back without writing any file. When all of them are done, write one file called compare.txt in the agent-tryout folder with one section per product. Use your web and file tools only, do not write or run a program, do not sign in, send, or buy anything, and do not touch any other folder.

While they run, Hermes lists the sub-agents above the box where you type, in both the desktop app and a terminal chat. The docs say sub-agents cannot write to memory, send messages, or schedule jobs, and a sub-agent that needs a command approved asks you in the same chat.

The Stop button ends only the current turn, and closing the desktop app does not stop the sub-agents either. Type /stop to end every sub-agent at once. The toolsets reference lists delegation in the default set for terminal chats, so if Hermes says it cannot hand work out, type /tools list to check that it is on.

What you should see compare.txt with one section per product and a page address beside each fact.

Safety note: on the local backend, every sub-agent works on your real computer with the same reach as your chat. Read each approval request, including the ones from sub-agents, and deny anything that reaches outside agent-tryout.

Cost note: set a spend limit at your provider before you start. The delegation docs say "a parallel batch of subagents typically burns the large majority of a run's total tokens," so this job can use many times the tokens of a normal one.

Sources

Research in this repo:

Official Hermes pages (checked 14 November 2026):

Next

Continue to Fixes.

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