First agency
Your First Real Task: opt out of data brokers, with your permission, and open a receipt on disk.
Set the approval gate (before you start)
Agency without a gate is a liability, so set this before the agent touches anything. Smart approvals ↗ let low-risk steps through and stop on uncertainty. Manual is slower and perfectly fine. Full auto-approve modes (sometimes labelled YOLO) are expert-only — leave them off for this course.
It lives in Settings → Safety → Approval Mode. Smart is the course default. Off is the one to avoid — that is the auto-approve mode.
- Set Approval Mode to Smart or Manual — never
Off. - Decide your rule now, so you are not deciding under pressure: approve steps inside the folder you are working in, deny anything that deletes or wanders outside it.
Know your cost (light)
Provider ↗ (OpenRouter, OpenCode, …) is where the key lives; model is capability and price. Free tiers still have limits. Before blaming the model, know what is already loaded in an empty session.
- Note which provider and model are currently selected — you will want to know which one did the work.
- Optional: open the context / usage view in Desktop and jot the two biggest slices (tools / skills / memory).
hermes prompt-size prints the fixed budget of a fresh session — note that tool schemas are the bigger half, and that it runs offline with no API call. @witcheer on X Put your agent to work
Pick one lane. Same job either way: install the skill, consent once, drain the queue, open the receipt. Official Hermes unbroker is US people-search — not this mission.
| You are | Skill | Law |
|---|---|---|
| Living in Geneva / Switzerland | unbroker-ge | nLPD / FADP, French letters |
| An EU or UK resident | unbroker-eu | GDPR Arts. 17 and 21 |
Read the approval prompt once. Approve official opt-out forms and writes inside the folder you named. Deny deletes, anything that wanders off, and any attempt to attach an identity document.
.txt draft you send from your own mailbox. Credit bureaus that demand an ID (CRIF, SCHUFA, Experian…) stay in the digest.hermes skills install github:deltaVgit/website-private/public/courses/open-harness/skills/unbroker-geunbroker-ge folder into this profile’s skills dir, e.g. %LOCALAPPDATA%\hermes\profiles\<profile>\skills\security\unbroker-ge\ on Windows. Not ~/.hermes/skills unless that is actually your install.hermes skills install github:deltaVgit/website-private/public/courses/open-harness/skills/unbroker-euunbroker-eu into that profile’s skills/security/ folder. New session after copy.- Install one skill (lane A or B). New session so Hermes loads it.
- Paste the matching line below. Give name, address, email, phone when it asks — once.
- Let it drain the queue. Approve searches and no-ID forms. Deny deletes and ID scans. Do not expect it to send email — copy drafts from
drafts/into your own inbox. - When it stops, open
unbroker-receipt.mdoffline. If the file is missing, the mission failed.
Use unbroker-ge. I consent. I live in Geneva.Use unbroker-eu. I consent. I live in the EU.| Lane A files | Lane B files |
|---|---|
| Robinson (paper) + YourOnlineChoices (ads) | YourOnlineChoices (ads) — first |
| local.ch, Moneyhouse, KünzlerBachmann, Intrum | Acxiom, Criteo, LiveRamp, Oracle Ads |
| Poste + Google results + Criteo/Acxiom/LiveRamp | Google results-about-you |
| D&B if a person page exists | D&B if found |
| CRIF / AZ Direct / Creditreform → digest (ID) | SCHUFA / Experian / CRIF-EU → digest (ID) |
If you still have time after the receipt exists: /learn on a file you own, or the Desktop HUD (“this” instead of describing a window). Neither replaces the footprint loop.
You deny a delete tool call. Next?
Pick one — the answer is revealed straight away.
Where the session ends and memory begins
Everything you just did lives in this session. Close it and the conversation is gone — the skill you installed stays, but what the agent understood about your work does not. That gap is the whole subject of Part II.
First, the session controls. Long chats fill the window, and these are how you manage it:
/context·/compress— see what is loaded, and shrink it/undo·/retry·/branch— conversation control/rollback+ checkpoints — files already changed (not the same as/undo)
None of that is memory, though. It only tidies the chat you are in. Everything Part II adds — notes it keeps, a folder of your own writing it can search, jobs that run while you sleep — starts from one idea: if you want it kept, it has to be written to a file. That is lesson 07.
Part I close
You now have: free model path, Desktop, soul, one messaging surface, vocabulary for Think → Act → Observe, and a file that proves the agent acted. Part II adds memory, vault, skills, security dials, and cron.
Other ways to deploy
Tool backends and host isolation
Shell tools use a terminal backend. On a dedicated host you may run tools more freely. On a personal machine, Docker ↗ (or a VPS) is the responsible default — not “later if you feel like it.”
- Dedicated host process — ok when the box is lab-only
- Docker backend — commands isolated from the host OS (required compromise on a daily-driver)
- SSH / VPS / cloud sandboxes — tools run on another machine you can wipe
- Own infra vs cloud API models — brain location is independent of tool backend; cloud brains still send chat content off-box
- Multi-machine ops — Desktop here, workers or gateway on a VPS
You are done when
Part II starts next: memory, vault, skills, security, and cron — the compounding harness.