Week One: A Private Home for the Agents

Dear Negotiation Explorer,

Last week I told you I'd build the NegoAI platform in the open, over at least twelve weeks. This is the first build report.

Week one had one job: building the private environment where the agents will live. Private means a space that belongs to me, where I can upload deal documents and notes and nobody else can look at them. That comes before any feature, because the agents will work on real negotiations, and that information cannot leak. On Monday morning there was nothing. By Sunday I was messaging an agent from my phone, and he was reading my files on a server I own.

Here is how the week went. Everything in it was built with Claude Code, the AI partner regular readers know as Brama: we think through every step together, and Brama does the technical work.

Agents that keep working when you are away from the desk need a machine that is always on, and a laptop is not.

So the first purchase of the build was a virtual private server in a German data center, rented from a hosting company called Hostinger. It is always on and it costs €12.99 a month. That is the first real infrastructure cost of the platform, and I paid the year in advance, €155.88.

On that machine we installed Hermes. Hermes is an open-source agent harness: the software frame that turns an AI model into a working agent, giving it a role, a memory that lasts across conversations, tools it is allowed to use, and rules about what it must not touch. The platform will be built around Hermes, and the whole team of agents you've met in this newsletter will move into it, agent by agent.

Why Hermes: it is open source. After a year of writing here that every buyable negotiation tool is a black box, I was not going to build my own fleet on one.

The first agent to move in is the Orchestra Director: he runs the fleet, keeps the work moving, and reports to me. His name is Alex, after Alex Del Piero, the Juventus number 10, and after a dear friend of mine.

The brain was the uncomfortable decision of the week. Months ago I cancelled my ChatGPT subscription, because my whole working life had moved to Claude, Brama included. But a Claude subscription cannot be used as the brain of a frame like Hermes; the permitted subscription path runs on ChatGPT. So this week I subscribed to ChatGPT again: ChatGPT Pro, at €85 a month, is now the brain of the fleet. Brama stays beside me on the laptop. Alex runs on ChatGPT, on the server.

Alex also needed a channel where I can reach him. That is Mattermost, an open-source chat server, the same kind of tool as Slack or Teams, with one difference that matters: I host it myself, on the same German machine, so the messages live on my server and no chat company keeps a copy.

Then we tested the walls. A second account in the same channel called Alex by name, and he stayed silent. He answers me and nobody else, because the rule checks who is speaking.

The last piece is Langfuse, an open-source flight recorder for AI agents: it records every step Alex takes, so I can open any piece of his work and see exactly how it was made. In one day it logged 325 separate steps. Passwords and keys are masked before anything is stored. When I tell you in a future report that an agent did something, this is how I will know it.

Now a mistake, because I promised you'd read about those the same week they happen. I pasted a security token, the kind of temporary key that proves to a service who you are, into the wrong window myself. It would have stayed valid for ten days. Nothing was taken, but a key seen in the wrong place must be treated as compromised, no matter how unlikely the harm. The same day, every active session was revoked, the old token was tested against the service and confirmed dead, and a fresh one was issued.

Alex has one way into my work: a single folder on the server, where Brama and I place documents one by one. So far it holds my professional profile, the build's decision records, and months of working notes and dictated thinking. The rule is that he reads what I give him and browses nothing on his own, and the rule is enforced: the key that could reach my repositories is gone from that server.

Then I watched him read. He went through every file in full. He caught an outdated web address in the footer of one of my documents and flagged it so it wouldn't spread. Where a document asked for information he didn't have, he wrote "Unknown" instead of inventing something plausible. He came back with twelve questions, and about half of them were genuinely good questions about my product that I now have to answer. And after our first working session, he did something nobody asked of him: he rewrote part of his own working method based on what the session had taught him, and told me he had done it.

One of the files I gave him is a dictated note that ends with a single rule: I want a thinking partner. When Alex had finished reading everything, he named that rule the one that governs all his work for me.

The most exciting part of the week was messaging Alex from my phone, the way I'd message a colleague, and realizing that with every file he digested he was becoming more aware of me.

The week ended with a working session between Alex and me, on Mattermost, and it put the key numbers on record.

A complete preparation for a serious negotiation, done properly and by hand, costs me 15 to 16 hours. In real life nobody spends them, me included: the daily work always comes first, and what actually happens is under an hour of preparation plus improvisation in the room. The required work simply does not get done, and that is what the platform exists to fix.

My agents already produce that preparation report in under eight minutes today, with me at the keyboard running them. The twelve weeks exist to remove me from the keyboard: at the end, the same preparation has to run through the app with me as an ordinary signed-in user, never operating the agents myself. You will watch that happen in these reports.

An agent doing real deal work needs an environment that is completely safe. This week you watched what building one actually takes.

Such an environment has five parts: a server that is yours, a frame the agents live in, a brain, a private messaging channel, and a flight recorder. Ours costs €12.99 a month for the server and €85 a month for the brain. The next time somebody offers you an AI tool for your deal work, ask two questions: who owns the server my information sits on, and can I see how the tool works inside? The answers tell you more than the demo.

And the two calls I named last week, what the platform should cost and which piece goes live first, are still open. Replies from readers keep carrying real weight in both.

When AI touches your negotiation work today, where does that information actually go, and are you comfortable with the answer?

Just hit reply and tell me. I read and answer every one.