Hermes Agent · user story
Taskaway is a small company in Rouen, France. We deploy AI agents for small businesses. Our own back office runs on five Hermes Agent profiles: each one has a single job, the machine enforces a few hard rules, and the founder decides what matters.
- Hermes Agent
- Five profiles
- By Otsoa Dubief, founder
One agent, one job.
Each agent is its own Hermes profile, with its own memory, skills and schedule. The founder only talks to one of them.
Lore · orchestrator
The founder's right hand, reachable on Telegram. She runs the morning briefing at 8 am and the evening one at 5 pm, goes through the inbox three times a day, hands work to the others and checks what comes back. Her briefings run on Claude Opus, her inbox rounds on Claude Sonnet; the four specialists run on Claude Sonnet.
Hari · sales
Researches prospects, drafts outreach and replies, and books the free audits.
Odei · communication
Website, SEO, the weekly video and social posts. For each post he prepares two versions; the founder picks one.
Ekaitza · deployments
Prepares audits, scopes each agent, sets it up for the client, runs the first week and hands it over.
Dirua · finance and admin
Quotes, invoices, deadlines and the paperwork of a small company.
The founder
Leads every client audit himself and decides on money, legal and security. Everything else reaches him as a short morning summary.
How they work together.
- One hand-off command. When Lore needs a specialist, she hands over the request with a single script. Each request opens a fresh conversation, signed by the sender, and the exchange is logged in a shared journal. No endless chat thread to re-read.
- A task register. Anything that is a deliverable becomes a task with an owner, a status and a proof of completion. Lore reviews it in the evening briefing.
- One calendar per agent. Every routine shows up in Google Calendar, next to the work each agent actually did. When the founder drops an event into an agent's calendar, it becomes a task for that agent.
- A weekly review. On Fridays, one page: what went out, what goes out next week, the numbers. The founder can stop anything with a word.
Rules the machine enforces, not instructions.
A prompt is a request; an agent can forget it. So the rules that matter live in code the agents cannot change.
- A guard plugin. It blocks sending email directly, reading secrets, and editing the guardrails themselves. Only the developer tooling can change it, never an agent.
- An isolated sender. Outgoing email goes through a separate system user. Agents can only drop messages in its queue; it checks daily caps and opt-outs before anything leaves.
- Reading the web closes the door. Once a conversation has read an outside web page, it cannot send anything on its own: a page can hide an instruction. The next conversation starts clean.
- A kill switch. Any agent can switch autonomy off with one file. Only the founder can switch it back on.
- Cards for what matters. Money, legal, security, and emails or invitations to outside people come to the founder as a Telegram card: accept or refuse.
What we learned.
One job per agent
A narrow role keeps the skills short and the mistakes easy to trace. When two agents could do the same thing, we write down who does it.
Code over prompts
The rules we care about most ended up as checks the machine runs, with tests. Instructions alone kept drifting.
A journal for everything
Every hand-off and every outgoing message is logged. When something goes wrong, we read what happened instead of guessing.
And in your business?
At our place, after a run-in period and automated checks, the team acts on its own. Every agent we deploy starts the same way: the agent proposes, you approve.
Every agent we deploy runs on Hermes Agent, on a machine and an AI account opened in the client's name. More on our French pages: the team at work, Hermes Agent explained in French, and our open guide hermes-agent-fr on GitHub.